erans

Also, since DeepSeek 4.1 Flash is already quite quantized, the same OpenRouter benchmark harness puts us at #9, but the first 10 places (out of about 30) are within less than 1% of each other - so that is much harder to set.

But here, again, we run on B200, the official weights with specific configurations that we have on vllm.

You are totally write. I wrote exactly about that. https://www.lunaroute.com/blog/the-model-is-not-the-product/

We will release our data soon. It's not just quantization, we are using NVFP4 from RadixLabs for GLM 5.3 but with our configuration and the OpenRouter harness for benchmark we are actually #4 in accuracy out of about 30+ providers. We run strictly on Blackwells (B200) on SGLang.

So its not just quantization, its the software, its version, its configuration, the exact hardware etc. They all matter.

I've been using DS 4.1 Flash for a while now as my main workhorse. It's amazing. It's fast and it is also a GREAT debugger - it might not always find the best solution but it will definitely find the problem quickly.

I'm also the CEO of a new company - LunaRoute - so if you want private, fixed cost, we run our own server in the US type of DeepSeek 4.1 Flash check out https://www.lunaroute.com

If you want a trial, hit the contact us and write that you saw this post.

We also have GLM 5.3 (with vision!) and GLM 5.3 Flash - all included.

For agents, if you are concerned about that, block access to "su" as it is interactive anyway. Not loading it into the memory will block the attack. If you are using AgentSH (https://www.agentsh.org) you can add a rule to block "su" and soon be able to block AF_ALG sockets if you want to further protect things.

Execution layer security must be deterministic. That's why we are working on AgentSH (https://www.agentsh.org) which is model, framework and harness agnostic.

The part that seems most important here is that npm install was enough.

Once the compromise point is preinstall, the usual "inspect after install" mindset breaks down. By then the payload has already had a chance to run.

That gets more interesting with agents / CI / ephemeral sandboxes, because short exposure windows are still enough when installs happen automatically and repeatedly.

Another thing I think is worth paying attention to: this payload did not just target secrets, it also targeted AI tooling config, and there is a real possibility that shell-profile tampering becomes a way to poison what the next coding assistant reads into context.

I work on AgentSH (https://www.agentsh.org), and we wrote up a longer take on that angle here:

https://www.canyonroad.ai/blog/the-install-was-the-attack/

It's great to see more such platform popping up. It's good for the ecosystem. We need more hosting options that are clear, secure and have the ability to help people run as many models as possible.

that's great. PagerDuty always felt so expensive and heavy!

That's awesome. It's always annoying using those 3rd party ones!

Summary data is in a sqlite database so its easily queryable by you or your friendly AI agent.

You can have a single proxy and have multiple team members use the same one so that you can easily track sessions, token usage, spend etc.

true that there is a some kind of a ceiling of what can or can't be done. But that ceiling is way up there. Also, there are enough examples and articles and code that allows enough combination to be made so that its good enough - and that is a very important bar.

There are A LOT of businesses (even big ones managing money and what not) that rely on spreadsheets to do so much. Could this have been an app/service/SaaS/whatever ? probably.

What if these orgs can (mostly) internally solidify some of these processes? what if they don't need an insanely expensive salesforce implementor that can add "custom logic" ?

A lot of times companies will replace "complex software" with half complex process!

What if they don't need Salesforce at all because they need a reasonable simple CRM and don't want to (or shouldn't) pay $10k/seat/year ?

There are still going to be very differentiating apps and services here and there, but as time move on these "technological" advantages will erode and with AI they erode way faster.

I also added a Mac version: https://selfhostllm.org/mac/ so you can know which models you can run on your Mac and get an estimated tokens/sec.

I would argue that due to the way MCP servers/tools are added to calls, there will be a pre-step that will figure out which MCPs are even relevant for a request prior to executing it.

Thanks :-) Fixed it. Don't drink and copy&paste ;-)

Most Thinkpads will work just fine out of the box. The hardware is great (I specifically like the keyboard and the little red point mouse).

The T series (and W series) are the work horses here with various screen sizes, disk options and RAM.

I personally use the X1 Carbon 3rd gen (X series) as it is comparable in specification to a Macbook Pro 13" (I have the one with the 16GB of RAM). It's a little costly but its worth while and will probably serve me for the next 2+ years.

Very interesting point. Stop talking and start doing :-)

I know this is too much, however we plan on adding the ability to inject sharing data directly to websites, that's why we asked for it in advance.

We would never do anything to miss use it and we felt that an optional permissions scenario was rather too complex.

How would you fill with an optional permissions scenario?

A little tool we developed as a weekend project to test and debug a bigger project we are working on.

We thought it would be useful to others.

Cool greeting cards!

It looks like a really interesting approach for live debugging...

We've been analysing sharing patterns for a while. We tend to see that Twitter shares correlate too much with the amount of posts per day which may indicate its more bot driven.

FB Likes and +1 are much harder to fake and seems to correlate more with actual human traffic.

From our experience +1 tends to be factored into Google search results so there is direct value in having +1 on your site and having your visitors who like the content click on it (not to mention it gets circulated inside Google Plus itself).

We are working to add more sites as we go along and we would be happy to share some information about patterns across sites and topics.

Thanks for the vote of confidence. We think its a great tool too. There is much to wait for with additional insights that we are thinking about.

Would love to get feedback from the community about which other types of information and analysis they would like to see.