Wheen
Wheen on Hacker News
One consideration is that Qwen3.5 has bigger/later versions that are already better at math. If it was indeed training on ML research, it's possible that what it is actually doing is applying lessons learned from the Qwen3.8 27B weights (which independently solved the problem) to Qwen3.5 0.8B.
I doubt it's that simple, but I could totally see an LLM being able to make that association and apply a fix from an improved LLM to a smaller one with the same architecture.
It's worth pointing out that this is a multi-trillion (probably) parameter model optimizing an 800 million parameter model. There is no indication that it can improve itself with this same methodology.
> [Anthropic] basically just need to prevent it from absolutely collapsing and hope margins hold up against open weight models.
There are so many variables involved that are outside their control. If GPU/RAM demand dries up (or supply increases), for example, local models are suddenly a threat in addition to hosted open weight models.
The quality there is already good enough. I've been 100% local for personal projects since Qwen3.8 Flash dropped. It's not as nice an experience as Claude I use for work, but it gets the job done and I fully control my data and can do whatever I want with it.
It's usually against the license to redistribute modified binaries. There are few pieces of software (I cannot think of any) that allow redistribution of modified binaries, but aren't open source.
As a result, in practice, the only example I can think of is cracks for pirated software.
I'm curious to see your source for these numbers. Your flow sounds closer to a 99th percentile one than a 25th. I find it hard to believe that a median developer is using 5-20 agents at a time.
> When you hear about the work Anthropic researchers are doing, they seem to be using hundreds or thousands of agents to solve problems.
There's a reason they're Anthropic researchers. The value proposition tanks when you actually have to foot the bill for the tokens you use.
Subagent flows can improve overall performance, but it generally also comes with disproportionately high token use. That gets worse with parallel subagents that are often doing the same things at the same time and cannot share a cache.
AA has also lost all credibility to me with all of their index changes being unannounced and following an open weight model getting close to the top spot. I gave them the benefit of the doubt for v4.2, but to do this again immediately after Kimi K3.1 got into the top 5 is a little too suspicious for me. For Opus and Astra to stay at about the same score while everything else drops feels like they are adjusting weights specifically in their favor.
Even in the optimistic case that this isn't some conspiracy, it highlights how arbitrary their scores are. I had trusted their previous ratings and their new index shows that even AA doesn't agree with them. What's to say this new index is any better and won't just change drastically again?
They literally don't even mention gaming revenue in earnings calls anymore. It's lumped in an "edge computing" category.
It's just slop. Tried to install it via flatpak. It has flatpak instructions in the readme. It does not ship the flatpak files it references in the readme.
It's a dependency of OpenClaw. Everything else aside, that's ~390k likely users minimum due to the auto-starring.
Though, I realize now the "hundreds of thousands" claim I was responding to is per week, while my guess is all time.
I think you're right, but in the wrong direction. I'd imagine it's closer to a million.
- It has 3.6 million weekly NPM downloads.
- 110k GH stars.
- It's 5th (and its fork is 6th and a dependent is 8th) in monthly Openrouter tokens. Add them all together and they get close to Claude Code numbers.
Not the person you're replying to, but judging by the emphasis on the cost of cached input tokens in the OP article, I'd guess it has to do with DeepSeek v4.1's KV cache efficiency. It uses <1000 bytes per token, so they're able to get 1M token context in under a GB.
Edit: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/...
> but where's the outcome dataset justifying this? how did they get that probability? what was the specific methodology of the tests? what variables did they account for?
It's just high school statistics: https://en.wikipedia.org/wiki/Confidence_interval
It's not a probability. It's essentially an assertion that if the test were run 100 times, the result would be within the interval 95 times.
Between all the posts fabricating scenarios to justify the AA score and the others trying to undermine AA, I'm getting strong astroturf vibes.
Either that, or the average poster on HN isn't nearly as critical as I had thought.
Have you posted the code to this proxy somewhere? I'm dabbling in custom harnesses and would love to learn from it.
Just a hunch, but it might be because of the 51B parameter n-gram embedding. At 125B, you'd expect ~16gigs for a 1-bit quant. Add 51gigs for the n-grams and you're not far off the actual size.
If that's true, it'd scale linearly with number of bits in the quant with an offset of about 51gigs. So Q4 should be a bit bigger than 82gigs, I'd guess in the 90s (as opposed to a ~280gig q4 if the whole 70gigs of the 1-bit quant scaled linearly).
Sorta, but you're off by one layer. You can store the 18B in VRAM and stream the rest from RAM. There's still a performance hit relative to storing it all in VRAM, but it's tolerable.
Generally, for local consumer use, these large MOE models are best for unified RAM systems like DGX Spark or Mac Studio.
> will rely heavily on public chargers
Public chargers =/= DC fast chargers (like Tesla superchargers).
I've already seen level 2 chargers that drop down from light posts in my city. Roll those out to every light post near apartment buildings and throw in some curb-side stalls if you need more capacity. Most people won't need to charge every night with 200/300mi+ range, so you won't need a 1:1 mapping of chargers to cars. This way, people charge passively overnight at stations that are cheaper to build.
As a former HBO dev, can confirm that the majority of test cases or random debugging messages are Game of Thrones quotes.
Sorry for the late reply.
Recently, mostly Python and Javascript
For me personally, just scrobbling.
At this point, that's the only thing stopping me from using it. Looks great, but I'm a stat freak and I need my scrobbling.
As someone who is (Probably a bit more than) a junior dev with no formal education or experience in a team, do you have any suggestions to write better code?
I feel like my code would be considered bad, but I have no idea what good or bad code looks like. Does anyone have some examples and comparisons between good and bad code?
I have a friend who claims his aunt (Who works of MS) said that it would cost $1k (This was literally the day of the announcement, too). He has been rather reliable in the past, but it should still be taken with a grain of salt, since I was given no context.
Airs shouldn't be that bad. IIRC, Linus runs one.