Hilarious sentence: "DeepSeek-V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks." (emphasis mine)

Kinda shows their team is more engineers than marketers.

But with that version number, it's not a "minor" update! It's a patchlevel update! -Engineer

An old-school (pre-LLMs?) engineer would recognize that DeepSeek is in no way doing semver, so your major/minor/patch glossary might not apply here at all :)

Strictly speaking, it just shows they aren't marketers. Bad copy doesn't tell us anything about their engineering skill.

It's not a bad copy, or at least that's not what the person you responded to is insinuating.

Deepseep-v4.1-Flash-1017 is a minor update. A major update is something like Deepseek-v5-Flash.

> It's not a bad copy, or at least that's not what the person you responded to is insinuating.

FWIW, GP didn't write "a bad copy" but "bad copy", referring to the marketing sentence quoted by OP, rather than the model that was just released (as I presume "a bad copy" sounds, maybe like it's called that because it's [again, presumably] distilled?). "Copy", especially in a marketing context, is a written message.

Strictly speaking, they didn't say anything about their engineering skill

Strictly speaking, it's the substantial improvements that tell us about their engineering skill.

Well, it's a bigger improvements in Deepseek 4.1 Flash and Deepseek 4.1 Flash 1017 than whatever Google was doing between Gemini 3.1 and Gemini 3.8.

It's not really contradictory, minor changes can have major effects.

So they meant a "small" update - "conservative", "inexpensive"...

An example when using a less frequent term actually reduces specificity.

No, they used minor correctly.

Will you also explain, or leave it like that?

Changing TREE(2) to TREE(3) is a small change on the input, but a... lets just call it big... change in the output.

Hoping it's minor relative to what's coming

That's one small step for a man, one giant leap for mankind.

It's about playing to two different audiences.

The CCP requires extensive certification for new major releases, to make sure a model doesn't do anything they don't want it to.

To get around this, Chinese labs only bump versions when they have to and usually claim it's a minor update.

Then on the other side, they're obviously trying to market to the tech crowd...

Source?

Everything in China is strictly regulated but loosely enforced. Just take that as a grain of salt

I happen to live in China now and lots of my friends work in frontier labs. I don't now there's a restriction for major updates of LLMs.

Western media doesn't have correct informtion source about China. This is because Western favors Web ecosystem and most inforamtion live in the Web, but in China everyone use Apps.

Drum Tower podcast (from the Economist, paywalled) about Chinese AI.

Would highly recommend it, it's IMO the only serious Western reporting about China.

> the only serious Western reporting about China.

as good as Gordon Chang.

I don't know who that is, but from a quick Google he doesn't exactly seem trustworthy?

then what makes you to believe highly politicalised publications like The Economist is "trustworthy"? please just feel free to google or AI search all those failed "crash" prediction made by The Economist on Chinese economy. Don't get too shocked to see there are dozens such predictions made by The Economist in the past 3 decades - it is beyond laughable. You get far better prediction results using any random number generator.

The problem here is quite often brainwashed people never realize that they have been brainwashed. Sometimes they trust, promote or even defend the perpetrators. You have my deep sympathy on this matter.

I am opposed to their political tilt altogether (I'm a socialist and there's literally a Lenin diss of the Economist!), and I disagree with their conclusions. That doesn't make the reporting itself useless, and they are the only ones going past a cliche caricature of China, even if they dress it in classic capitalist delusion.

There is no such thing as an unpoliticized publication, but one can still read, taking into account their biases. The point is to extract the information while discarding the ideological interpretation. In fact, this exercise is often easier when the ideological tilt is as obvious as the capitalist liberals from the Economist.

Forget about their predictions or their moralizing, but they have actual Chinese reporters talking to actual Chinese people on the ground instead of doing the usual Western thing of white MBAs treating China like some alien civilizational threat.

It's a bit refreshing to read updates from an AI lab without bullshit doomerism and overwhelming sense of self-importance from shit like OpenAI and Claude.

Also, DS is an excellent model.

I'm at 1 trying to get to 10. I move to 2. Minor update. 100% improvement towards goal.

These are not mutually exclusive.

That is how exponential functions work. Improve every aspect of a system with 1% and you may get easily 50% more performance.

Minor updates can provide substantial improvements sometimes, it really depends on the specifics.

Remember when we were dumbfounded why our service is slow despite having implemented quite robust caching mechanism. Turns out simply the flag that enables it was hardcoded to false. So it was minor update (changed flag to true), but provided substantial speed up.

Quiet confidence backed by provenance :).

They changed how it works substantially. I think it should be 4.2.

To me this implies they continued the training run for 4.1 and it improved. A 4.2 would imply they changed something about how it’s trained/how inference works, and a 5.0 would be a major update (who knows what would have changed).

This is just my headcanon though.

It passed my threshold for not caring to try new models back with DS4, I use it for literally everything. Announcements of new models in the series always feel curious, because it doesn't really feel like there's much left to fix that would improve on what's already there for any task I know of.

It's a bit like obsessing over patch releases of some software after a major upgrade already solved the original problem long ago.

I have quite the opposite experience: I can do any concrete coding change with any model, but the intelligence limit just doesn't go through more complex debugging examples.

And then when I try a smart model, like Astra medium and it solves it (I had a firmware bug in my new laptop where I couldn't charge through USB-C) on the first try (altough with a repo already having lots of information about the hardware/firmware/debugging programs/traces), I realize how much time I'm losing on the few examples that the less intelligent models are just not able to debug and solve.

Very cool!

Is your laptop something like framework which has open source firmware?

I(and many other people) have a similar issue with usb-c charging not working consistently. I've been meaning to try one of these new smart models to help me debug the issue.

ASUS Zephyrus, but it was able to reverse engineer the ports and I was able to download the firware from the controller itself, write an emulator that emulates running the firmware on multiple controllers with different instruction sets (the USB Power Delivery controller is RISC-V) and the communication between them. Also it found out how to read the communication buffer that was on the devices for debugging, which helped it understand the USB data packets that were doing the handshaking and voltage/current negotiation.

I had to reboot my laptop a few times though until I got to this, but it was a fun experience anyways.

At this point open source isn't as much advantage as long as the firmware is not encrypted :)

Wow, that is impressive. I've been meaning to do this by hand to fix the issue on my Lenovo IdeaPad for a while but haven't gotten around to it.

I'll try handing it off to one of these new frontier models.

Also they can work in completely new domains, like Astra and Opus 5.5 starting to be able to use something like Blender. Currently they still often do horrible topology and basic UV mistakes, and the animations are really bad, but who knows where they’ll be in 5 years.

5 years? Look at the leaps since last November. Since the Spring, even.

I don’t think these are completely new domains in training. Like I’ve taken up offers to help code up AI training material just to see what they’re up to and get a sneak peek at how the sausage is made, no idea who these front companies are working for of course, but they pay for all kinds of fringe cases and Blender wouldn’t even be considered fringe in what I saw at the beginning of the year. (I was only there to observe, never submitted anything, so I didn’t betray you guys by helping train your replacements.)

I’ve heard about this through vfx artists

I was really hoping this would be solved since openai said that gpt-6 could paint brushstroke by brushstroke but it’s still shit

Speed and price. You should try the 4.1 Flash on one of the fast providers. It's pretty crazy how fast this model works, and it's great for iterative software development because you can steer it fast to where you want to go.

https://www.lithosai.com/

https://www.coralbricks.ai/

How are these high-speed providers close to matching the API rates from deepseek?

Seems that V4.1 Flash is quite cheap to serve. Especially how it implements the cache, meaning the cache pricing can be very low depending on how you implement it in your inference engine.

Depends on your use case. But if you do software development with a harness such as oh-my-pi, your cache rates are probably somewhere above 98%. Coralbricks has free cache reads, meaning it's much cheaper than the Deepseek API rates.

Lithosai is more expensive especially in the ultra tier. It's if you want extremely snappy and fast responses, which is important e.g. for a personal assistant where every second counts.

If you generate a lot of output, you're better off with the Deepseek API rates.

Ive never needed an assistant pre or post ai to reply to me in 100ms at a clip of 500 words per sec. See you in 15 min, Im making some cofeee.

same here, but surely the trends will follow every other profit-seeking web app ever and distill down to a competition over attention.

Where does one even get decent ZDR provider? I'm wary of sending sensitive data but not at a scale where I would build my own machine.

Scaleway have DS4. I don't think they call it ZDR, but their framework is very strong https://www.scaleway.com/en/docs/generative-apis/reference-c...

You just search around, or heck, just ask Google for inference providers that are ZDR and whatever else you want. Just make sure to read their terms/privacy and review the security center. You will have to request access if you are wanting to get into the deeper details.

Outside of major players like Amazon/Google which have compatible offerings available, I believe together.ai, fireworks.ai, and deepinfra.com are, as best as I can tell, US-based (if that matters in your situation) SOC2/HIPPA/ZDR. Each of them offer serverless model hosting with caching, which is going to be the cheapest overall. Most of them also offer private AI deployment which is more expensive but your entirely own instance. Some of them more heavily quantize their models (deepinfra) so take that into your workload considerations. I included deepinfra because they tend to have some the new models sooner than others or have models the others don’t (like mimo). IMO, together > fireworks > deepinfra.

However, you’ll have to keep in mind that for true ZDR you give up caching which increases your inference costs, so it depends on how “Zero” you need the zero to be and on what timeframe. If you just don’t want your data being trained on, normal use is fine imo. If you are working on something actually sensitive (like, because of NDA agreements and such), then they might not be the right choice and you may be looking at more expensive offerings or building a machine.

One note: you have to pay attention to the source of the inference. Many (similar to open router) offer models that are pass-throughs to a providers API, but they are pretty clearly marked. Just make sure the model is marked as serverless when evaluating new ones. This is where a proxy comes in handy so you can prevent your LLM from accessing a different models from the specific ones you’ve vetted.

I like privatemode.ai for better than ZDR, encrypted attested hardware (in EU)

Ollama Cloud have ZDR.

> improve on what's already there for any task I know of

What about "good governance"?

Does it have zero hallucinations? Does it reflect thoroughly on every single idea it uses, verifying its solidity?

definitely could use less thinking time and lower token usage, which is what this bump is focused on

Yeah, but for some reason people think they need SOTA to help them work on their side/indie projects.

Like if we needed PhD intelligence for productivity tools and SaaS platforms, why weren't we hiring PhDs to build things before.

99.99% of people do not need SOTA ai.

Sota is not about intelligence. It's about behavior. Glm 5.3 flash is dumber that got 5.5 or opus 4.8 but better to use because it's more modern. it can manage subagents, handle longer running tasks, be steered, etc. it's better at doing the thing. Same thing with opus 5.5. fable 5.1 is probably smarter but opus 5.5 is better

Same. I'm just happy sticking with them and am consistently impressed. Cost is so good and kind of keeps you honest.

Above GLM 5.3, nice! That is a very capable model. I do wish that we will see some smaller "flash" models again though - running this one locally is tough (very costly).

For comparison, close to Opus 5. Far behind Sol 6.1 and Opus 5.5.

Probably a pretty decent model for 90% of the usecases!

I've been doing a lot of hobby work with GLM 5.3 Flash and I can count the amount of times that I need a stronger model on two hands.

I have access to unlimited tokens at works and Opus is great but also often overkill if you know exactly what you need.

I use Opus for planning, through API on medium thinking, it doesn't cost anything significant and the rest can be happily done with deepseek or glm flash. Using sol 6.1 for review loops at the end. All running on pi.

A very complex task tends to cost like 10$ max. It's comparable to a subscription with enormous flexibility.

It's a steal.

I'm using DeepSeek a lot, and I'm curious how you get a task to go to ten dollars with it? Or are you mixing Opus and Sol tokens to the same task? I have very hard time to go over 5 dollars with DeepSeek and that is already a weeklong project where I share the same context.

Of course I do certain things like use omp and do a compact picture snapshot after 256k tokens automatically. My tasks rarely go over a dollar...

That's the total cost of a task across all models. DeepSeek going above 2-3 dollars is actually rare because of the very low cache costs.

I also used to plan with more expensive models like you did, but now I just do planning with DeepSeek too and don't see much difference. I use the omp plan mode and iterate on the text until I'm happy with it. Then clear context, give the plan to DeepSeek again for implementation.

If it's mission-critical code, do a security review with GLM 5.3.

Otherwise go manually through the code and ask for changes if needed.

Works really well for me.

I'm curious why you use Opus for planning and Sol for reviewing, instead of the same for both? Do you find they are better for those tasks?

Opus 5.5 is, generally speaking, a much better model than Sol 6.1. But I usually see that sol can discover much more edge cases than opus, and it's much cheaper (as in, subscription usage as well as token costs).

Honestly, I don't know what people see in Sol 6.1. Opus 5.5 is so much more inteligent, as in proactive, avoiding stupid mistakes, where it contradicts itself, ignores the context. Sol 6.1 has a gold-fish memory, it's as annoying for me to work with as Opus 5 was.

You can override the context length of Sol to 1M

Without this change I hated Sol and was burning tokens way faster due to all the forced compacting.

Since making this (obscure/hidden) change I enjoy Codex much more.

The work I do cannot fit in 256k so it was just relearning the same things over and over again while making zero progress. AFAIK the 1M burns tokens at twice the rate but I use Sol on medium as the work I'm doing now needs more memory than it needs intelligence and it's working great. I use Opus 5.5 for the stuff that needs stronger intelligence.

Fair. I bumped the context size, will see how it goes.

That's not actually the problem. It forgets the instructions from 15 minutes ago, without context compression involved in-between.

That has not been my experience after bumping it up to 1M. I do run a lot of agents in parallel ~10-16 and I actively babysit them so perhaps I'm just used to reminding the agents with things I've already told them. But I don't notice much of a difference between Sol 6.1 1M and Opus 5.5 1M

I'd rank Sol 6.1 only slightly behind Opus 5.5 in intelligence and both significantly ahead of anything else I've used (though cheaper/faster models are totally fine for a lot of tasks). My only real annoyance is 6.1 Sol is one of the slowest models I've used, which does make Opus 5.5 more pleasant to use.

> and both significantly ahead of anything else I've used

Really? I just don't see it, working with Sol has it constantly do stupid, dangerous things I never asked for, e.g. "I’m sorry. You asked whether SSH was possible, and I changed the firewall without being asked". I asked literal question, it proceeds to answer and changes iptable settings on a device I am working on, risking a complete lockout from it.

It is genuinely dangerous and has to be looked at all the time. Fable, Astra and Opus 5.5, but also Deepseek 4.1 are ahead of it. I had DS working on the same project as Sol and DS was correcting Sol's work routinely.

I never understood the hype with Claude

I tried Claude legitimately for a few months, on max

That was during the great anthropic compute/gaslight crisis (ahh good times..), Claude then enthusiastically bricked my MacBook and I had to get gpt to unfuck it

By next year we will likely see GLM 5.3 level performance in Qwen 4 MoE models that can run on most higher end consumer systems.

I don't think there is a reason to care about any of it, just get a few dollars on a few openrouter or similar platforms and just use the models for now, wait for cheaper models that are good enough for you to become available for your hardware.

Sure you might never see frontier performance on consumer hardware but do you really need that though? Does anyone? And if you ever do just get a 30-day sub for that work, and cancel it.

Yeah performance at a given compute budget will likely continue to improve.

We are running Qwen 3.8 27B FP8 with 4x 5060 16gb at the office for a small development team. It is doing pretty well for implementing backend and frontend tasks, and is already useful. In 6 and 12 months from how I expect it to get even better. Some planning is still done with a stronger model, but even that is maybe not strictly necessary.

Qwen 3.8 wasn't that nice in my experience Qwen series has lagged a lot but I think Qwen 4 would be better they cleaned up the architecture.

And we are likely to see many companies start building 27-32B MoE models custom built for use cases now that Rubin is going to make training costs much lower.

I expect training costs to get under 100k USD within a year for small models with another 100k maybe for RL.

I think just having all your own data from claude and stuff on your system that you can feed these models and have companies offer great base 27B models will be how it works.

Cant you run it on a pair of DGX spark? Sure it's expensive now, but at the initial retail price that was what, 8k in hardware? Not exactly unachievable

No, this is not a small model. I believe 4x Sparks is possible but 4.1 is pretty big for a flash model.

MiMo 2.6 Flash is the most capable model you can currently run on 2x Sparks without resorting to a quantised version of something.

I meant quantized. You can get an EXL3 quantized version thats almost the same performance.

My experience running quantised models is that they look the same at first but are all noticeably worse once the context size gets beyond anything trivial.

I've tried both 2bpw and 3bpw quants for QWEN3.8 27b, and while i do notice a difference in terms of higher rates of getting stuck in loops, they are both reasonably usable.

Are they as good as opus 5.5? No.

Are they about on par with opus 4.6? Yes.

The long run tasks i run seem to be about the same level of quality, though there are obvious differences in model styles, token budgets, etc....

Very keen to try this out tonight. What coding harness do people like to use with these models?

You can stick with Claude Code if you're already using it. DeepSeek has an Anthropic-compatible endpoint, so you just need to set ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic and ANTHROPIC_AUTH_TOKEN to your DeepSeek API key.

Pi is very good with open weight models (proprietary ones too)

Pi if you're a minimalist

OpenCode if you want a good, opinionated setup

DeepSeek Harness works very well, also used Codewhale which is not too shabby

Thank you deepseek, we are eating good today.

Looks very impressive especially considering the price and speed. Let's wait for AA to confirm how well their benchmarks hold up.

You trust AA since the sol fiasco a month or two ago?

What did I miss?

They updated their methodology when the new openai model didn't rate higher than the old one, or something like that.

I do t think it was a fiasco. I thought the updates were good. They should do more work on benchmarks. There has been good work on testing benchmarks for example to ensure they are accurate, and as issues arise they should change the benchmarks. The point of AA and Arena is trying to understand how to think about models.

Why not just try it yourself?

It costs money to run serious tests. We're way past the time when you could get an idea of the new model in a couple of tricky prompts. Now I need to compare a whole resulting project to get a clear view. If someone else proves the benchmarks don't hold up in reality, then it saves us further tests.

I'd never heard of the BabyVision benchmark they do so well on. Pretty interesting stuff:

https://arxiv.org/abs/2601.06521

Is it coming to OpenCode et al?

Presumably. The current opencode slug for the model is just 'deepseek-flash' suggesting they will just update that one going forward, not keeping versions in parallel.

and do we know which version runs on opencode at the moment (on the free and on the paid version), is there a way to find out?

The only way I'm aware of is to check their twitter feed. They moved to Deepseek v4.1 flash in September

sure, but which version are they running. they could be running an older revision

Just yesterday there was just one public revision of 4.1, and they listed they were using 'Deepseek v4.1 flash'. I assume they will update their website if they update the model version (and use the version in the name).

Yes, they've made public announcements (and changed the version string) on every previous Deepseek update

That announcement arrived 2 hours ago (on x), so now they have.

It's better than GLM 5.3 and Kimi K3 and way cheaper.

that's honestly insane. it's like the go-to for when you don't want to use opus 5.5/sol 6.1

The go-to is a local model.

Limp-to, maybe.

With Strata, running GLM 5.3 is possible.

$$$

Those are crazy numbers for a Flash model. Hopefully they can scale what they are doing to 4.1 Pro.

So this is the reason for the crash a few days ago.

What crash?

Deepseek servers went unavailable for 5 hours. They compensated that by rewarding the previous week usage as credit though. Pretty nice.

Ai/tech stock crash

beating K3 and GLM-5.3 is kinda insane

Another big win for opensource AI.

BTW who would have thought China would be the champion of freedom and opensource in the AI arena, while US is fighting for corporate/government control.

Well, at least mostly everyone outside the US.

Anyone that has observed China since the 90s.

There are different types of freedom and China's bet is that as long as they provide enough economic freedom the public will be ok with somewhat restricted social and political one.

Yeah, I had a somewhat odd conversation with a Chinese acquaintance last year when she got to talking about how America sees China. Her point was that, "How am I not free? I can do whatever I want, go where I want, buy what I want. I can talk to you and say whatever I want; nobody is here listening to me."

The fact that she can't call Xi "Winnie the Pooh" on Douyin wasn't something that occurred to her, or maybe that she was aware of at all. I didn't ask. I don't like to get into these discussions when I'm in China.

There is no such thing as “free speech” anywhere. It’s all a spectrum and the Americans in particular are indoctrinated. It’s a red herring to distract from the real social issues. People in America have some of the worst healthcare in the world. But hey at least we have “free speech”.

Fixing healthcare or economic policies is typically an act of congress type of thing - not even an amendment. Large popular support - and it can be fixed. Fixing freedom of speech or the ability to vote at all - once it is gone - usually takes a price payed in blood. Sometimes rivers of it.

Socialized healthcare does have large popular support, though. Lots of things have large popular support in the US and still don't get done.

It's a red herring. There is no free speech in US. Here's your free speech: https://www.facebook.com/katievbentetv/videos/-youre-embarra...

Citizens of Palm Beach county cannot voice their opposition to how their tax dollars are being spent. They rightly want that money to be invested locally, instead the County uses it to invest $1B in Israeli bonds, something that people widely oppose.

They first said that it's a budget meeting and Israeli bonds are a "policy matter". Then they started arresting people for speaking out against it. Then they started arresting anyone who applauded.

"Free speech" -> I spit on American "free speech". It's the most insidious and dishonest thing. At least the CCP doesn't PRETEND to give people rights that they simply don't have.

wish hn allowed reposts just to reiterate this point.

the U.S allows "free speech" if it's inconsequential.

I can't call AI "AI" without being termed an "enemy" by the President.

> The fact that she can't call Xi "Winnie the Pooh" on Douyin

Actually she can do this. There's plenty of such videos on Douyin ... If you clicked such videos frequently, algo will recommend a lot of them to you.

The worst case of uploading such videos is an aaccount ban (but very rare). So yes the freedom of speech here is fewer than in the U.S., but it's not like you'll be put into jail if you say sth about the president.

How do you even compare the amount or level of freedom of speech? One either have it or not have it! Stop pretending.

That economic freedom is a fantasy fabricated so that the king can rob people fortune created by the themselves in the name of greater good.

That's an image China wants the rest of the world to believe. Towards inside it still fits the typical stereotype.

China is only doing this openly because it benefits them to price dump the US.

"a minor update" they say before showing off benchmarks that display a substantial improvement.

they must've been using this since 2-3 days when it became free again in OpenCode, it's noticeable smarter, beats GLM

Presumably they registered, but that doesn't mean the model passed through export review. It just means that some bureaucrat now has the opportunity to check the product characteristics they reported and decide whether that kind of thing should be allowed for export or not. (Though with the weights out in the open, that's maybe a bit pointless.)

We're in a crisis

Censored, no surprise

Those benchmarks can't possibly be real, right? Right? <insert padme 4-panel meme here>

At this rate the frontier labs will be Anthropic and DeepSeek, with OpenAI struggling to keep up.

OpenAI's models -- and even Google's Argon -- have particular niches where they shine. For GPT-6/6.1, it's 3D modeling; Astra is hands-down the best at making .step files, Blender files, etc. For Argon, it's natural language writing and everything that hooks up to Google, e.g. search and maps-based itinerary planning.

Anthropic and DeepSeek (+ GLM) are definitely solving coding, though!

They're all quite different tools.

Artificial Analysis recently released a dedicated 3D benchmark, where Astra lags behind both Opus and Sonnet 5.5 (and yes, open weight models seem to be quite far behind).

https://artificialanalysis.ai/articles/aa-3d

Interesting!

For me it's large files in Blender (in some cases >400MB), mostly, where Astra excels. I gave it a look, but I don't know if the benchmark encompasses that sort of thing.

.step files are simpler pieces of straightforward code, so it makes sense that Opus and Sonnet 5.5 do very well...

its unfortunate they didn't include deepseek in the initial batch of benchmarks

I believe these are the last sets of super-high-end models available to public (open or closed).. these companies probably will now start to withhold top-of-the-line models for themselves, governments, and their super-high-paying partners only.

News from a near future:

Here's How Distill APT Used McKinsey's Private Access To Top Anthropic Model To Help The Chinese Keep Up With The Us Top IA Labs Despite Access Restrictions

That hasn't been the way to bet. DeepSeek and the other Chinese firms show no signs of slowing down, so if OAI and Anthropic halt releases and "pace the frontier" they'll just fall behind. (And there's a chance that high-end AI will become fully commoditized, which suits DeepSeek just fine.)

Besides, Fable 5.5 is reportedly releasing within the next few days, probably on Tuesday...

China just announced tightened export controls on AI models, for what it's worth.

I really, REALLY hope this is a joke.

I wish it was…

https://www.reuters.com/world/asia-pacific/china-proposes-ad...

They’ve technically got expedited approval for open weight models I think, but it still needs approval. And frontier level stuff gets extra scrutiny

We have the weights, we have the papers, nothing preventing development on this foundation.

Except compute

Do you think we are near the theoretical maximums of how much model we cram into 20B?

I don't consider myself qualified to make that call, though I believe there's plenty of room for improvement. But I don't mean inference compute, which is fairly accessible for well-off enthusiasts even now. Training is a lot more expensive though.

Back in the day we had seti@home, prime95 and fold@home. We will make it work.

When models are in the hundreds of gigabyte range, there are not enough nodes to train, because most people dont have 100k in AI hardware. Also there is no training method for machines scattered around the world as far as I know.

idk, sol 6.1 is both cheap and capable. i've been using it exclusively, i don't need to care about usage, and it's really smart.

yes maybe opus 5.5 is stronger and flash is cheaper, but 6.1 strikes a good balance imo.

OpenAI models don’t come close to competing with deepseek in cost quality ratio. I have a $20 plus plan and I run out of my 5h limit in literal minutes, every tick of the harness is like 10% of my 5h limit.

You can in contrast get far more done with $20 worth of deepseek api spend. No comparison.

maybe, but sol 6.1 is far more capable than deepseek, and i love deepseek

i have a feeling that Sol is just Astra low, but with lowest possible priority on the compute.

I believe it's Astra quantized. It was briefly named astra-lite.

Try adding a new comment

I'm not sure if this is relevant or not, but just in case it is: please don't do any automated posting to HN. It's against the rules: https://news.ycombinator.com/newsguidelines.html.

I wonder how you would be able to reliably rely on a model like that for agentic work, even for implementation only after careful planning, how are you making sure you are not wasting tokens reviewing and correcting it over and over again ?

Yeah, I have the same question. I'd like to incorporate cheap & fast models like DS into my workflow to reduce my reliance on OpenAI/Anthropic, but after trying several such models and having to have a frontier model review/fix/redo everything multiple times per task, I gave up and simply have the frontier models (Sol/Opus) do everything. All of this fits pretty comfortably in my monthly subscription.

Maybe the equation changes if I exceed my subscription usage, and need to fill the gap with something else. But it seems like a huge waste of your time to go out of your way to use a cheap model if you have the tokens to spare in your subscription.

Here are some pelicans. The "max" one took over 11 minutes: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

Definitely an improvement on deepseek-v4-flash-0731: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...

Wow. Is DS the only model that consistently puts one leg behind the frame? I don’t even remember Opus ever doing that.

Some of the most recent Claudes and GPTs have done that too, and Gemini has done it for a while.

You should start asking for animated three.js pelicans too, the 2D pelicans seem to have saturated.

Faithful to the original theme. Apropos. +1 on adding the prompt as long as it’s affordable and easy enough Simon, and not too distracting/noisy/time consuming all considered!

here you go! Results are trickling in https://fenderog.com/art-crit/page.html?p=surfing-pelican

The best bicycle is "low".

This is in "high" mode .. 40 seconds only: https://ibb.co/5XTRKNJv

Looks much better.

This one has ample space in its beak.

Does DeepSeek provide the inference code along with the weights? I read here recently that for models where only the weights are released, providers have to implement which could be different than what the labs themselves use.

There's basic reference inference code released with the weights, yes.

https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017/...

> providers have to implement which could be different than what the labs themselves use.

I'd imagine (most) providers just take implementation by vLLM/SGLang.

At this point I am way more excited by DeepSeek releases than anything from OpenAI or Anthropic.

The speed and cost/intelligence ratio is just outrageous.

Oh mince, mon immense meurtre des humains congois dans le fruit

oh is “Flash” a cool sounding word again since Adobe abandoned it?

Honestly given how good the open weight models are getting I don’t see a viable business model for OpenAI and Anthropic anymore. The profit is in the compute, not the model. OpenAI and Anthropic spent all their money renting compute from others so yes they can sell lightweight model inference but they have a massive structural disadvantage now as others can always sell good-enough inference much cheaper.

There’s also zero stickiness to models and most companies now have their own internal routing utilities so devs and users can just keep the same API endpoint and move between models with a single line of code swap.

Unless something changes with the big labs and fast, they look pretty f*cked.

People will pay for more intelligent models. Astra and Fable are more intelligent than these open weight models and can solve problems that are much more difficult, and that's extremely valuable. It's not any more complicated than that.

> It's not any more complicated than that.

It is more complicated than that when the valuations and business plans are contingent on those companies AI having complete dominance at every level.

If companies only use frontier models for occasional difficult tasks, not only does that affect volume, it basically elminates stickiness and incentivizes companies to instead of using their products continuosly, find ways to move from the expensive stuff to the cheaper stuff.

Yes but that’s a tiny portion of AI use cases. Maybe 1-2% max. The big labs need a lot more than that to survive.

You’re assuming other model creators can’t build these mega models. Evidence is they’ve simply decided it’s a fools game to do so and they’re focused on where the use cases are.

Fable is behind Opus (which is also much cheaper) at this point in benchmarking.

I don’t know what I’m talking about but it seems to me OpenAI is heavily emphasizing productization of LLMs. Dots, ChatGPT Work, scheduled tasks, in-app projects etc.

Being able to directly distill their huge frontier models into Luna/Haiku-level models puts them a step ahead still.

And yet, nearly all of enterprise money spent on AI tokens in the US is going to either OpenAI or Anthropic.

That trajectory appears to be changing. Virtually every company I see is shifting more and more to open weight models. Anthropic’s slowing growth rate is already spooking investors and that shift is just getting started.

That's why they are constantly fearmongering and trying to get other models outlawed. If nothing else, under the pretense of national security.

i hate dario's politics, but i still find claude's eco system easier to use. Design/Artifacts is very convenient for people like me who has zero UI/UX expertise, and claude code's remote session works seamlessly on ios with a well-design client interface while you have to deal with ssh and typing in cli with your iphone keyboard if you are using opencode/omp etc

Agree with you on model quality, but Anthropic has done a nice job developing stickiness. The connector ecosystem, integrations with desktop apps, browser use, etc. are all pretty compelling for a typical business use scenario. It's expensive, but also a "safe choice" akin to buying IBM or Microsoft back in the day. My workplace (not just developers) is deeply invested in the Claude ecosystem & it would be pretty disruptive to pivot it.

How far could they jack up the price before you would consider switching?

Most corporate will not be using deepseek, haven't heard of it.

There is either no correlation or a negative correlation between product utility for price and Enterprise and Government purchasing decisions so OpenAI and Anthropic should do just fine.

They made very bad bets on scaling, which has cost ~10B scale in wasted capital allocation, but it is manageable. Their purchase commitments for future compute capacity are absurd nonsense, but all of that is just shtick for retail investors, and they will reshuffle the deck as needed to adjust.

Anyone using DS as CLI coding agent which harness do you recommend? Is there anything as good as CC/Codex?

I use deepseek via fireworks using Claude code.

When I type deepseek on my CLI it launches Claude with my fireworks API key. Works very well and feels almost free.

I really like pi, though it does require some time in setting it up. But it has greatly reduced my token spend. The relevant Deepseek Harness https://www.deepseek.com/en/harness/ I hear is also good and popular.

As a Claude Code main, I've really enjoyed working with DS4F in omp. https://omp.sh/

I am a huge DeepSeek fan, but it cannot come close to Opus 5.5 on UI design. A $20 Claude subscription that I use some months and cancel the others along with Deepseek API is great.

What would be the best local harness to run this? It feels there are 1000s now.

Exactly, there is no ”best.”

Claude code 2.1.16x (released 1+ year ago or so) is quite ok, dynamic workflows off. Do not use the latest version, they burn a shitload of tokens and are slow.

at this point there are 100s of benchmarks.

so AI labs can just pick which to show

- how good is it at reversing a game exe file?

I use Opus 5.5 and swarms of about 50 deepseek models to reverse engineer old videos games. The scale is roughly 600k loc generated from pure asm, simply by knowing what the old compiler is. Truly staggering stuff.

Could you please share your methodology?

See here https://news.ycombinator.com/item?id=50135410

Registered just to ask you.... 50 instances? Which harness? How? Why? Can you post a video demonstration?

Caveat: I get a budget of free ai tokens at work, as I work for an open source AI inference provider.

My goal is to reverse engineer an entire gamecube game -- I'm pushing an existing community decomp that was stalled to the finish line, solo, with lots of AI help.

This work was directly inspired by https://decomp.dev/ -- which has just shut down a few days ago because of, ironically... people like me who enjoy decompiling things with AI, so I'm still weighing / grappling with the ethics of how to release my work, it's murky waters at best.

I use Opus 5.5 as the orchestrator, as well pi harness because it's somewhat lightweight. But for the really hard functions I don't use deepseek I use opus managing other opus agents.

I'm working to release something soon as proof that I can do a 100% decompilation job, but the main idea is that when decompiling a gamecube game your sources of truths are individual functions or files, so you can build off that primitive.

Each ai gets a worktree, and a shared giant context file that's basically a record of every experiment that has happened. They're assigned tasks to essentially do as much as they possible can for each individual funciton. After about an hour claude uses a script to merge the work that's actually correct (verified by tooling that checks if the c goes to asm) and then there are usually 3 rounds of reviews by independent ais, because there will always be some slight mistakes. The techniques used as well as the mistakes the models make are recorded in the giant ledger. There's a bunch of random one off scripts such as stuff to clean up, and tools to make sure the agents don't accidentally google stuff / delete stuff / use too much space.

Claude steers the ship via ssh on my strong linux tower at home so it keeps my laptop nice and cool and at this point I just check in from time to time, I am basically a manager, and Claude has it's own workers.

Last point -- I don't know any c besides the basics, I've only been "coding" python / ts for a couple of years, but even then with the extreme strength of models the future is managing agentic swarms, so I'm leaning in as much as I can. I really see this as the future of large scale projects. For some people it's a mockery of decompilation, but, my goal is the finish line, not the journey.

Web search: $0.75 per 1,000 calls

That's way cheaper than openai and other web search providers. I have no doubt that AI demand justifies all the compute money big tech is spending now, but I also think there are so many over valued VC funded companies that are going to crash in the next few years.

Here's a small benchmark that compares the new checkout with the previous one:

https://slopgrid.com/deepseek-41-horses

The new model is amazing. Had it build a full Angular Material & CDK components showcase via DeepSeek Harness and it easily did the best job of any recent models I've seen.

Where are the Unsloth quants for DS-4.1-flash?

Still need 2 or 3 days to properly quant this. But it is really good model to run if you have 64GiB or larger RAM on Mac. On M5 Ultra, you get ~3400 tok/s for prompt processing and ~50 tok/s for token generation, quite usable state.

Forgive the ignorance, but how are you running a 763B parameter model on 64GB of RAM?

SSD weights streaming. The actual weights is 552B, then engram.

which engine?

By “quantitizing” it

The smallest quant is still over 100GB, so I don't know how that fits in 64GB RAM. It would have to be partly streamed from disk.

Once read, I think from G Huntley, "pick a model and stick with it," they have been making that very easy, daily driver since V4.

A negative datapoint - We have been evaluated prior Deepseek Flash models (both the original 4.1 flash and the smaller 4.0 flash) and they have been overthinkers. Our workload is well defined but moderately difficult coding tasks and among open weight models GLM5.3 flash performs best with Qwen3.8 flash next a close second. DeepSeek 4.1 flash fails at lower thinking levels and spends a huge amount of tokens (an order of magnitude more) at the highest thinking levels where it is able to successfully complete the tasks. I will evaluate this newest release since they specifically claim "the model reduces task execution time and output length", but keeping my hopes low.

What is your workload?

v4 was a worse overthinker. 4.1 mostly fixed that. It still prefers to plan everything before making a single file change.

If I'm impatient I switch to Luna for a single turn just to get it started and let DS continue.

This model is why OpenAI and Anthropic want regulatory hurdles. We need more diversity like this, not less. A diverse ecosystem is the only long term solution to complex problems that I know of and the challenges that the AI world is presenting is no different.

DeepSeek is simply amazing.

Their strategy is regulatory capture.

Through fear. Specifically making the American public so scared of AI that they hand over regulatory keys to the only “experts” that can keep them safe. This is a centuries old political strategy used by some of the worst regimes including apartheid South Africa where I grew up. It’s horrifically effective.

I disagree that they’ve invented some “strategy” for this.

1. They, and many other experts, have been saying the same thing since well before the current paradigm existed. 2. There is nothing about regulation that is required to give them an advantage; it’s a fallacy to suggest “regulation = advantage” 3. It would be preferable to argue the facts (is it or is it not dangerous? Why or why not?) than conspire as to why they’re stating what they and plenty of others have been saying for decades

You are right that they didn’t invent it. It’s a pretty well-established method for cornering a market.

1. All those who have been arguing this - many of them non-technical - are part of the same ideological movement.

2. Technically, there isn’t. In practice, majority of the time regulations benefitted those who wrote them.

3. Not sure if this is a criticism of the comment you are replying to or of the p-doom crowd. We would all appreciate facts stated instead of science fiction tropes repeated.

Look, “a lot of people say this” is a Trumpian fallacy. A lot of people say a lot of crazy things. It says nothing about the factuality about what they say. How many are this lot? Who are they? What do they know? Why are they incapable of expressing it without so much emotion? How come a big portion are non-technical?

Hinton and Bengio, and previously let's say Stephen Hawking are not part of the same ideological movement.

I'm pretty sure Hinton and Bengio get a lot of news coverage and are heavily over-credited for saying these things, unlike LeCun, Ng, Sutton, or Schmidhuber.

In fact, their core ideology actually stemmed from Yudkowsky's arguments.

They certainly didn't invent this strategy, but that's because it already existed. Regulatory capture is nothing new.

> 1. They, and many other experts, have been saying the same thing since well before the current paradigm existed.

Some experts sure. "Some experts" say lots of things.

However, this is actually much to their advantage. If you were a frontier AI company, a legitimate concern to use as justification for your regulatory capture is exactly what you want. That is, after all, generally how regulatory capture gains public approval.

AI safety was also why OpenAI had to charge per-token for GPT-3. It must've been such a struggle for them to make that difficult decision. I'm sure Sam Altman spent a lot of nights crying to sleep over that one.

It's good that we never got open weights for a GPT-3 class model, right?

> 2. There is nothing about regulation that is required to give them an advantage; it’s a fallacy to suggest “regulation = advantage”

It's difficult to outline the number of different ways that regulatory capture can and will benefit OpenAI and Anthropic.

If we limit things to "Fable"-class models, then nobody competing with Anthropic/OpenAI is ever allowed to outdo them. If you literally legally aren't allowed to make better models, you can never charge the "premium" of having the current top-tier model. You have to compete by being cheaper than Anthropic instead, which is a lot less enticing from a standpoint of trying to become profitable. Take a look at the per-token pricing of Fable real quick. Yeah. Being able to charge a premium for the model that does the cybersecurity thing is literally their most important asset.

If Anthropic really is making a profit now as rumored, then they need regulatory capture as fast as they can possibly get it. Once Anthropic and OpenAI prove that AI is a viable investment, entities with a lot of capital will suddenly become even more interested in re-allocating it towards taking a slice of this pie.

> 3. It would be preferable to argue the facts (is it or is it not dangerous? Why or why not?) than conspire as to why they’re stating what they and plenty of others have been saying for decades

OK. On the factual end of things, AI safety experts have been raising the alarms for over a decade and so far essentially none of their predictions have come true.

The rationalists, the main group pushing the AI safety rhetoric that dominates the Western AI sphere and is virtually absent in China, which appear to be going entirely off the deep end at this point in time, put together their best minds to produce AI 2027, which of course like most predictions started to fall apart a bit before hitting the one year mark. Predicting the future is hard.

I have made this point before but I will make it again: I think AI safety experts deserve at least as much skepticism as Ed Zitron in their claims. They've been making similar dubious claims about AI capabilities and an incoming singularity for a long time now, and so far the best we can say is that maybe it will eventually come true.

Maybe. Or maybe you've read too much goddamn science fiction.

They’re for sure losing this.

They seriously underestimate how ready and willing Americans are to use Chinese providers. There’s a ton of spite

Last time China pulled this shit it shook portfolios for a week

> Some experts sure. "Some experts" say lots of things.

I get this is Hacker News so we have to hate AI safety, but you should read more on this topic rather than blithely casting the x-risk viewpoint as fringe. The last hard data we have on the question is a late-2024 AI Impacts survey [1] of 1,580 authors at top AI conferences, who were offered $50 for answering a survey on the topic of the future of AI progress. The median respondent assigned a 5–10% chance (!) to "extremely bad (e.g. human extinction)" outcomes.

> On the factual end of things, AI safety experts have been raising the alarms for over a decade and so far essentially none of their predictions have come true.

Devastating. When I wrote my AI alignment PhD thesis [2] in 2022, and predicted a 70% chance existential risk by 2072 (which I later came down to about 1/3 on), you're right that that prediction hasn't come true yet.

While I agree many specific worries didn't materialize (like AI being like a "literal genie" it's hard to specify goals to), many key worries have (Hugging Face rogue AI swarm scans as cribbed from rationalist AI scenario writing).

[1] https://aiimpacts.org/wp-content/uploads/2026/09/ESPAI2024.p...

[2] https://arxiv.org/abs/2206.11831, top of page 4

You have to realize by now that outside of academia nobody is even paying any attention to people who are predicting existential risk that far out. For worse or even worse, AI 2027 is the AI safety conversation we're having today, and I regard AI 2027 as essentially Anthropic fanfiction, except usually fanfiction is written better.

I suppose, for you, an extinction question that didn't mention LLMs is evidence that Anthropic should be given regulatory capture of LLM development!

What paradigm?

You're saying regulatory capture gives you no benefit by just refuting an absolute with 0 context. There's nothing to even debate here. Ok... most people disagree?

Yeah I agree with point 3... Stop engaging in vageries and say something of substance with examples, so people can actually have a conversation with you without us doing all the legwork.

I keep seeing this assertion. What regulations have they proposed that would have prevented this new model? I thought they want to pace the frontier, which this model is not.

You are taking their words literally, rather than looking at the side effects of any effort to "pace the frontier".

Literally zero, but that never stopped a good conspiracy theory, especially when half of HN hates the big labs because of fear for their own economic future.

OpenAI and Anthropic aren’t even slightly worried about DeepSeek.

Having looked at the sentiment around DeepSeek's models around this thread as well as my own experiences with it I'm going to need a little more to back up that assertion about OpenAI and Anthropic not being at least a little bit worried about DeepSeek...

I’ve been listening to this rhetoric about the open models eating O/A’s lunch for like 18 months, meanwhile their revenue is up what, like 50x?

May this kind of worry find me…

The reason is quite simple, American corporations know they cannot depend on models developed in China; not because they're bad or lower quality models, but because they know their government can ban these models any time they want. It's an implicit assumption that every company knows to be true.

It's quite different here in EU. I just gave a presentation in an AI conference in Paris about moving from Anthropic to open-weight models, how it looks like and what kind of savings you'll be looking at. The room was packed, and after discussing with a bunch of VPs, staff engineers, and CTOs in the conference, every single one of them are either looking at or using DeepSeek now.

It's quite hard to have the valuation of Anthropic if they are limited to US.

This doesn’t make any sense either, the use of these models is on the rise throughout corporate America, it’s just that they’re augmenting the frontier models, not replacing them.

You said it correctly: augmenting, not as a main solution, because of what I said above.

People like open models because they are free/cheap. It's not anything deeper than that. Most people will hop in whatever van is handing out free candy.

It’s about capability.

The model makers are bricking the models with their system prompts, and even the tools are tamper proof.

You can only use it their way. They think we’re stupid.

And their way is the shit way, the babyproofed way, while they talk to you about their unfucked models hacking out of their labs and hacking into other labs, and reminding you you have zero rights.

Fuck them twice, then once more… This is about compute. Who owns compute. China is saying fuck you all we can scale things bigger than you can imagine. And they can! They have so much more energy! Theyre closer to commercial fusion than us too. Meanwhile our guys are trying to… regulate? Lol

China won’t have more compute than the US for least five years. They’re highly constrained.

Five years is a long time

When they surpass us, they will surpass us hard, in a way that is nearly impossible for us to compete with them.

Training and serving LLM's is massively parallel. Your bottleneck is either in energy or chips. China will catch up on chips fully decades before we catch up on energy. Arguably, this is already happening: here in the USA, we have lots of chips, but we are seeing our electricity bills skyrocket as more datacenters come online. We are powering those datacenters with natural gas turbines, not solar, wind and nuclear. Right now, to purchase one of those, you are on a 5-year waitlist. There's your timeline!

China is invested, from a regulatory, cultural, workforce, strategic, and technological standpoint, in an energy abundant future. We are not invested -- and in fact, actively sabotaging ourselves -- on all of those fronts.

Once they acquire chips, we lose, basically permanently. Look at what Huawei is doing recently and what is on their roadmap to understand how fast this catching-up is taking place. I'm sure that David Sacks would say that our innovation motor is pulling their copycat caboose along, as our monopolist tech companies steam ahead towards the frontier. I am not so sure. They are already ahead of us in many technologies. Speaking of technologies, they're building natural gas turbines in house, starting in 2025.

This reads like someone who's worldview comes from social media comment sections. China is still an emerging economy, don't confuse emergence with expansion.

Let’s see what happens

I agree that in the long run, China will likely surpass us in AI capacity. Our refusal to build out renewable capacity is a ridiculous self own in the long run.

> I keep seeing this assertion. What regulations have they proposed that would have prevented this new model?

> Literally zero, but that never stopped a good conspiracy theory

Are you living under a rock and never read any of Dario's blog posts?

Let me give you an example, here: https://darioamodei.com/post/policy-on-the-ai-exponential

Section "1. *Regulation* and public safety" (emphasis mine), then under it:

> AI companies that develop advanced AI models must have strong security standards that protect their model weights

You can't "protect model weights" if the weights are openly available, ergo this is an explicit call for regulation to make open weights illegal.

You're the one extrapolating that cherry-picked sentence to mean "open weights should be illegal", but that extrapolation doesn't make any sense, and it's hardly "explicit".

Presumably you'd want the AI companies to protect their weights to keep them from falling into the hands of bad actors. But the US can't make open weights illegal anywhere except the US, which does very little to keep them out of the hands of bad actors (most of whom are outside of the US).

Go read Dario's screed on GLM, or his other one one on how your toaster could be part of a distributed open model that self replicates. Every second blog post is some doom scare about the dire consequences of open models. Go look at the new ToU [0] which are now pages and pages of things you can't do with Anthropic models. Check how they will ban your account if you are persistently mean to Claude now.

He's completely transparently campaigning for regulation. It must be very frustrating to him that the Trump admin is completely non-responsive. He probably thought Anthropic would have the whole market stitched up and water tight by now.

[0] https://www.anthropic.com/legal/aup

Dario has not (and will not) say he wants to ban Chinese models outright. But:

1. He's made many posts about how he thinks "dangerous" models are bad (https://www.anthropic.com/news/position-open-weights-models)

2. He writes posts proving Chinese open weight models are "dangerous" (https://insightsintegration.com/anthropic-warns-of-advanced-...)

3. He talks about about how he wants stronger govt regulation of AI (https://www-cdn.abcnews.com/Business/exclusive-anthropic-ceo...), and explicitly mentions he does not trust China at all and doesn't want them to have Mythos.

Do the math: 1 + 2 + 3 = he wants to regulate/ban the Chinese open weights.

Of course, Anthropic and OpenAI will then write posts about how their own models are dangerous - but that's okay, they're the good guys.

Honestly, open weight models are dangerous, and I haven't seen a single convincing argument against that.

Not being able to audit my own software and configurations for vulnerabilities is dangerous. Consolidating power over the majority of Americans' jobs is insanely dangerous. Need I continue?

There are no choices here which do not carry risk. I haven't seen a single convincing argument against these points.

I'd argue that OpenAI and Anthropic are far more dangerous and I also haven't seen a single convincing argument against that.

Other things that are dangerous: kitchen knives, cars, household chemicals, not tying your shoelaces, political parties, water. Danger is a normal part of life. Rather than regulate or ban anything dangerous, what we normally do is assess specific risks, and only put a limited amount of regulation in place where it's absolutely needed to prevent the most harm.

We also put responsibility on humans to act responsibly. In theory, it's absolutely insane to allow human beings the ability to walk into a street with automobiles, because obviously they will be killed! But we don't ban them from walking there, and we don't ban cars, because it's pretty uncommon for people to walk directly in front of a moving car, despite there being obvious danger.

Ultimately the biggest risk of AI isn't the AI. It's what human beings do using the AI. Just like with hammers, knives, drills, chemicals, cars... we regulate the human's actions and create laws to direct their use, rather than regulate the items themselves.

Knives are dangerous.

> I haven't seen a single convincing argument against that.

When HuggingFace was in the midst of defending itself from OpenAI's automated hacking spree, they were blocked from using frontier models by Anthropic and OpenAI for "safety" reasons. HF had to resort to an open model to defend itself - and DeepSeek distinguished itself in that role: the second fact from that episode OpenAI would have you soon forget.

100%

Existential risk can be another big reason

This can be a reason too. There is an another reason, they want some govt involvement because good relationship with govt, so govt will help them with public money, when the bubble bursts.

PSA Deepseek train on your session if you use their endpoint and/or harness.

I'm a communist so I'm fine, but you may not be, so be aware

PSA OpenAI&Anthropic train on your session if you use their endpoint and/or harness. I'm an American! so I'm fine, but you may not be, so be aware

Have security companies considered that these open models could be used as sleeper agents even when they're deployed outside of China? Is that a serious attack vector or more like science fiction?

There are theoretical attacks involving malicious weights. But the risk profile is so outrageously in favor of them not doing it that you'd essentially have to be an anti-China conspiracy theorist to find it plausible, let alone likely.

They would most likely be bad sleepers. Here an illustration: https://youtu.be/HBURlcNpfoo

This is a great model. I know we are all looking for the next best thing, but I'd be pretty happy running this for awhile.

And Christmas comes early this year.

DeepSeek v4.1 Flash was a beast.

It was already cheaper and more powerful than Haiku 5.5 which was released like a full month after it... by a trillion dollar american "frontier" company.

Can't wait to see what this one has in store.

Getting a lot of errors from OpenRouter's endpoint "Failed to get response from the AI model; retried 5 times (total retry wait time: 60.71 seconds) Last error: 429 Provider return error"

It works when it works, but will probably be flaky AF for a few days.

Works well with Copilot:

    COPILOT_PROVIDER_API_KEY=<OR key> \
    COPILOT_PROVIDER_BASE_URL=https://openrouter.ai/api/v1 \
    COPILOT_MODEL=deepseek/deepseek-v4.1-flash-1017 \
    COPILOT_PROVIDER_MAX_PROMPT_TOKENS=1048576 \
    COPILOT_PROVIDER_MAX_OUTPUT_TOKENS=700000 \
    COPILOT_OFFLINE=true \
    copilot

700000 because there's no detail yet on the safe completion limit for most harnesses (OR lists 1M for both which makes no sense). COPILOT_OFFLINE enables all calls except to OR (no GH servers).

I haven't put this into test as much yet.

As a Luna6 proponent, this has a potential as it looks slightly more intelligent and at the similar price point.

Mario one-shotted by deepseek-v4.1-flsh-1017 (in high mode, not even max) with sound and programmatically generated assets in under 5 minutes: https://mario-dsv4-1-flash-1017.pagey.site/

Have you tested the game? Jumping on the pipe after the koopa and goomba kills mario

I can't wait for a smaller moe dinstillation from this excellent model

Why no Jinja-format chat template?

I essentially only use DeepSeek. Almost always the flash variant. I've found that I don't really need frontier models if I'm doing most of the thinking myself and just using the agent to output code quickly.

I am very happy with what DS has been able to provide me in terms of output. Also being off peak hours saves a lot on the cost side of things. I would recommend it to anyone that is dabbling with an idea and intends to use AI because DS will produce something that can provide a solid proof of concept.

I asked opus 5.5 to run 3 "hard" questions against haiku 5.5, sonnet 5.5, and this latest deepseek 1017 as worker agents running claude code... and deepseek won.

Not only did it come in significantly cheaper even than Haiku, but it was more correct than Sonnet.

This is great! I am gonna get coding!!!

Crazy how you can have AI almost doing everything now while you sleep.

I'm glad the open source frontier is advancing, but man, with Opus 5.5 I feel no need to look elsewhere. Such a good, cost-effective model for full-time work on the $100/mo plan.

I have both platform.deepseek.com (hooray off-peak pricing) and the $100/m Claude subscription and boy do I find myself mostly using DS-flash pretty much every time.

There's something to be said for the speed with which it gets word done - I take fewer "it's compiling" breaks, and lose my train of thought less frequently. What's crazy is that the speedy work it does is pretty much always correct - which is a reason why I don't feel the need to spin up claude-code and "ask the expert".

This speed & good-work combo is the DS signature IMO

Fwiw I find gpt Sol on medium or low effort to also be faster than opus to the point of helping me not check out, but is quite smarter than dsv4 flash

OpenAI subscriptions are shit tier. Deepseek does 90% of what I need to do for pennies. There is no comparison. And I’m talking about working on relatively hard domain like gpu kernel optimisation and HPC

https://github.com/computerex/dlgo

If you want your 5h usage to runout in one call feel free to use OpenAI

You read like a shill for Deepseek interests, and that includes your comment history.

Imagine everyone online is a bot.

Or just us Europeans who are so sick of American corporations milking every penny with their LLM models and preventing us from using the full model capabilities.

Now we have: a ton of competition in different inference providers, very low costs, and speeds reaching 1k tok/s in certain providers.

We can just get a box of compute even in Germany, and run the exact same model here without anybody stopping us.

Are we using the same Thing? Sol 6.1 has run <30 t/s for me for weeks.

What would you prefer, 30t/s 50k output tokens needed for solution, or 300t/s, 500k output tokens needed for solution?

This is why "tokens per second" alone is a deeply flawed comparator.

This is also why discussions of Chinese models focus on tokens per second- they struggle to compete on intelligence so they dramatically increase effort levels to compensate, meaning they require massive total output tokens.

For example, flash 1017 isn't benchmarked yet, but 0731 required 15.4M tokens to complete the artificial analysis benchmark, and Sol 6.1 high required just 2.8M.

Sol 6.1 uses 2.8M at 64 tok/s = 12h 9m

Flash 0731 reported 149 token/s at 15.4M needed = 28h 42m

And in my example in the first line, scenario A (30 t/s) completes 6 minutes faster than scenario B, despite 10X slower inference.

I would prefer it work and solve quicker. Opus 5.5 has a reasonable pace to me. Sol does not. I understand efficiency I just don't like using it right now because it has literally hung or failed to complete tasks a dozen times in the last week due to inference capacity.

> What would you prefer, 30t/s 50k output tokens needed for solution, or 300t/s, 500k output tokens needed for solution?

The best way to answer this question: [Pareto frontier or cost/per task curve](https://api-docs.deepseek.com/img/v4.1_261017_perf_cost_en.p...)

img from https://api-docs.deepseek.com/news/news261017/

I certainly don't think the self-reported single benchmark scores are the best way to answer the question. I waited for the Artifical Analysis run (which itself is still mediocre, but at least it's an order of magnitude better), and predictably the results tell a different story.

Sol 6.1 High: 50 intelligence, $0.32 per task, 4.4 min per task, 13k output tokens per task

Sol 6.1 Medium: 48 intelligence, $0.21 per task, 2.9 min per task, 8k output tokens per task

Deepseek 4.1 1017: 45 intelligence, $0.26 per task, 4.8 min per task, 89k output tokens per task

If you look at the actual pareto frontier across the full benchmark, Sol 6.1 remains firmly on the pareto line, while deepseek remains firmly under it, just as it was before https://i.imgur.com/N72p5NF.png

This result is exactly as I expected, and seems to run counter to the effusive praise of this thread. Sol 6.1 Medium is producing higher quality output at a lower price with nearly 10X less output tokens! And this is not surprising.

And finally, Mimo V2.6 Pro continues to look like a stunning release, clearly smarter, cheaper and faster than deepseek v4.1 1017

"Sol 6.1 Medium is producing higher quality output ..." -- When it is actually up and working. Even last night I still received constant capacity messages halting codex over a dozen times and I gave up and shut the workflows down.

This wasn't my experience. GPT 6.1 Sol is ~6-7x more expensive for similar intelligence. My experience matches [this](https://api-docs.deepseek.com/img/v4.1_261017_perf_cost_en.p...)

What harness do you use for deepseek?

opencode and deepseek harness GUI (dsh-tui when I need to)

After using it for 24h in DSH, this is a very good model and much better than the previous version. Cancelling my Codex sub when 4.1 Pro is released, or maybe before then.

Tried it. Little whale is getting smarter. Or, as DeepSeek might say: "Hmm. Wait! Hmm hmm."

Am I misreading or is this cheaper and better at software engineering than Opus 5.5?

Tried it and it works surprisingly good.

It's great at coding, it's superfast and it's a technological marvel, but talks just like Opus 5 and it's just as insufferable. A shame. Can see myself switching in some months, but I'm not going back _there_ as long as I can avoid it.

What a shame, I feel like this is a curse most of the Chinese labs have. Every other model is pretty good and stands out, then the next one is just a clone of Claude that does better on benchmarks and can oneshot Minecraft 2 or whatever but is just worse to use.

The real moat deepseek has is the tiny KV cache. This makes it basically the only serious choice for large local models. I hope to be proven wrong!

I feel like I’m an outlier based on all these comments. This is still working through my personal benchmarks, but it’s spending 30% more time and performing worse than the previous 4.1 flash did.

Benchmark is historical PR reviews. Models are put in a sandbox with a git and db snapshot as it existed at the time. PRs are a selection of changes some bugs (major and minor), some with things that aren’t bugs that have tripped up models in the past and some no findings that they should approve and move along.

4.1 flash already wasn’t that strong at this task, but the speed edged out v4 pro for a spot in the roster.

But does it still have the crazy hallucination problems that V4.1-flash has? I would guess so.