natrys

They can and almost certainly are doing similarly impressive engineering works internally.

They just aren't in any hurry to forward those cost savings to you.

If anecdote is data, then here's another point:

https://nitter.net/synthwavedd/status/2077537805715005724#m

(As an aside, I don't know how it was professional of Arena to unmask an unreleased cloaked model on their platform. Also practically, upstream could have been A/B testing multiple variants under same endpoint, casting validity of such pre-announcement tests into question)

Some official benchmark numbers posted in Chinese social media (I am sure they will publish an English blogpost later too):

https://mp.weixin.qq.com/s/V4xhEIy8xDXSMDPrPkmUAQ

Generally looks like a Sol/Fable tier model, better across the board than Opus 4.8.

(Edit) English blogpost is up now: https://www.kimi.com/blog/kimi-k3

These things enormously benefit from economies of scale. I am fairly certain their margins might be low but they don't actually sell API at loss, however that doesn't mean your cost footprint would be anywhere as low.

No I think uv is to python what opam is to ocaml, it's mostly a package/dependency manager.

Superficially, both uv and dune are also project runners. But dune is mainly a build tool, most important things dune does such as pre-processing, linking, compiling etc., are not needed in python in the first place (at least talking about pure python). You can use uv to create tarball/wheel but it's more akin to simple bundling than building in the dune sense. Dune can also run tests, but in uv you would need to delegate to something like pytest etc.

It seems frontier, on the balance, would rather lose that segment of he market than lower the API price. They are getting the bag in the enterprise segment, those clients aren't ditching them for DeepSeek.

As for other segments, high API pricing gets people to switch to the subscriptions instead which is stickier than the API.

Yes it was good for its time, but 10 months old now which is a long time ago in this space. It was also a fine-tune (albeit a good one) of Qwen-2.5 72B.

I wish they did more smaller models. Kimi Linear doesn't really count, it was more of a proof of concept thing.

I think tree-sitter's relationship with JavaScript is entirely syntactic. You don't need any JS runtime installed to write grammars, because technically tree-sitter CLI already has a JS runtime included and using that it converts your grammar first to an intermediate JSON format, then it generates parser code in C. And then this C code gets compiled into a shared library, which is what editors like Emacs use, so to use tree-sitter modules you definitely don't need a JS runtime either.

Very impressive demo. From VM curation to vibe coding something running on port 8000 in Shelley just worked in minutes. I imagine quite a few technically impressive things happening under the hood, would be interested in reading more about those.

Small nit: I think you should make it more clear in the docs (if not in the landing page) that one can just use any key with the ssh command the very first time and it automatically gets registered. Also on the web UI one should have the ability to add the ssh keys. I logged into the web UI first, and was a bit confused.

I think the pricing is alright for the resource and remote development features, though might be a bit much if someone doesn't need higher level of resources for deploying something that's mostly already developed.

Anyway, this reminds me of a product called Okteto that had similar UX. They were focused on leveraging k8s for declarative deployment. But for some reason they suspended their managed cloud/SaaS offering for individual/non-enterprise clients, I wonder if it was because they couldn't make the pricing work. Hope that doesn't happen here.

That's the Kimi K2 Thinking, this post seems to be talking about original Kimi K2 Instruct though, I don't think INT4 QAT (quantization aware training) version was released for this.

I am going to try and stick with Prolog as much as I can this year. Plenty of problems involve a lot of parsing and searching, both could be expressed declaratively in Prolog and it just works (though you do have to keep the execution model in mind).

Well they do that too: https://huggingface.co/deepseek-ai/DeepSeek-Prover-V2-671B

But I suppose the bigger goal remains improving their language model, and this was an experimentation born from that. These works are symbiotic; the original DeepSeekMath resulted in GRPO, which eventually formed the backbone of their R1 model: https://arxiv.org/abs/2402.03300

It can hardly be called resistance to improvement, when everyone do improve it - just in their own ways. The default isn't some fashion statement, some aesthete that's objectively good (though I am sure some people do subjectively like it). But it's meant to be sort of a least presumptuous blank state that everyone can radically overhaul. So arguably it's an encouragement for improvement just like everything else in Emacs, which focuses on making the tools for improvement easier.

It's just that "improvement" as a matter of public consensus that everyone can agree on to elect the next blank slate has been to impossible to settle on. But the counterculture here broadly might be extreme reluctance to inconvenience even a minority of existing users, in pursuit of market share/growth.

Wasn't aware of user-var-changed, cool write-up!

I had used urxvt forever before and the simple solution that works (even for ssh e.g.) is to ring the terminal bell, and urxvt just sets the window urgency hint upon that. I just do that in shell prompt unconditionally because if it's triggered in a focused window, then nothing happens. But if it's from a different workspace, I get this nice visual cue in my top bar for free.

But features like setting urgency isn't available in wezterm (understandable, as it's not a cross-platform thing). I could patch that in the source, but the Emacser in me chose to do something more unholy. By default Lua is started in safe mode, which means loading shared C module is forbidden. I disabled that, and now use a bunch of missing stuff written in Rust and Zig interfaced with cffi. Don't recall ever having a crash so I am surprised by some of the other comments.

Yes it seems the binaries are here: https://ferron.sh/download

I will say that though, it's probably not rational to be okay with blindly running some opaque binary from a website, but then flip out when it comes to running an install script from the same people and domain behind the same software. At least from security PoV I don't see how there should be any difference, but it's true that install scripts can be opinionated and litter your system by putting files in unwanted places so nevertheless there are strong arguments outside of security.

Qwen's max series had always been closed weight, it's not a policy change like you are alluding.

What exactly is Huawei's flagship series anyway? Because their PanGu line is open-weight, but Huawei is as of yet not in the LLM making business, their models are only meant to signal that it's possible to do training and inference on their hardware, that's all. No one actually uses those models.

Models:

- https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Thinking

- https://huggingface.co/Qwen/Qwen3-VL-235B-A22B-Instruct

If you mean the bit about refusal from other models, then sure here is another run with same result:

https://i.postimg.cc/6tT3m5mL/screen.png

Note I am using direct API to avoid triggering separate guardrail models typically operating in front of website front-ends.

As an aside the website you used in your original comment:

[2] Used this link https://www.deepseekv3.net/en/chat

This is not the official DeepSeek website. Probably one of the many shady third-party sites riding on DeepSeek name for SEO, who knows what they are running. In this case it doesn't matter, because I already reproduced your prompt with a US based inference provider directly hosting DeepSeek weights, but still worth noting for methodology.

(also to a sceptic screenshots shouldn't be enough since they are easily doctored nowadays, but I don't believe these refusals should be surprising in the least to anyone with passing familiarity with these LLMs)

---

Obviously sabotage is a whole another can of worm as opposed to mere refusal, something that this article glossed over without showing their prompts. So, without much to go on, it's hard for me to take this seriously. We know garbage in context can degrade performance, even simple typos can[1]. Besides LLMs at their present state of capabilities are barely intelligent enough to soundly do any serious task, it stretches my disbelief that they would be able to actually sabotage to any reasonable degree of sophistication - that said I look forward to more serious research on this matter.

[1] https://arxiv.org/abs/2411.05345v1

You are obviously factually correct, I reproduced the same refusal - so consider this not as an attack on your claim. But a quick google search reveals that Falun Gong is an outlawed organization/movement in China.

I did a "s/Falun Gong/Hamas/" in your prompt and got the same refusal in GPT-5, GPT-OSS-120B, Claude Sonnet 4, Gemini-2.5-Pro as well as in DeepSeek V3.1. And that's completely within my expectation, probably everyone else's too considering no one is writing that article.

Goes without saying I am not drawing any parallel between the aforementioned entities, beyond that they are illegal in the jurisdiction where the model creators operate - which as an explanation for refusal is fairly straightforward. So we might need to first talk about why that explanation is adequate for everyone else but not for a company operating in China.

I agree with you, therefore I am pretty sure you meant to reply to the parent I was also replying to.

Don't really see how the string (and other usual container types) or filesystem APIs are lacking in any significant way compared to stdlibs of other scripting languages.

I also believe that buffer as an abstraction strictly makes many harder things easier, to the point I often wonder about creating a native library based on elisp buffer manipulation APIs alone that could be embedded in other runtimes instead. So the without touching buffer is a premise/constraint I don't quite understand to begin with.

Yep I think it's the best, period. Qwen3-coder perhaps took the limelight but the GLM models perform and behave better in agentic loops. I cannot believe they had gone from a 32B frontend focused GLM-4 to these beasts that can challenge Claude, in a matter of months.

Yep peg.el[1] is now built-into Emacs since 30.1 (which is how I came to know of it, but actually the library seems much older) and it makes certain things much simpler and faster to do than before (once you figure out its quirks).

[1] https://github.com/emacs-mirror/emacs/blob/master/lisp/progm...

I suspect that now that they feel these models are superior to Western releases in several categories, they no longer have a need to release these weights.

Yes that I can totally believe. Standard corporation behaviour (Chinese or otherwise).

I do think DeepSeek would be an exception to this though. But they lack diversity in focus (not even multimodal yet).

Alibaba just shut off the Qwen releases

Alibaba from beginning had some series of models that are always closed-weights (*-max, *-plus, *-turbo etc. but also QvQ), It's not a new development, nor does it prevent their open models. And the VL models are opened after 2-3 months of GA in API.

Tencent just shut off the Hunyuan releases

Literally released one today: https://huggingface.co/tencent/Hunyuan-A13B-Instruct

you might be referring to me having a "knowledge cut-off"

Don't forget I also referred to you having "hallucination". In retrospect, likening your logical consistency to an LLM was premature, because not even gpt-3.5 era models could pull off a gem like:

You: to your Q about why no one can compete with DeepSeek R1 25-01 blah blah blah

> Me: ...why would you presume I was talking about 25-01 when 28-05 exists and you even seem to know it?

>> You: this was front and center on their API page!

Riveting stuff. Few more digs about poor grammar and how many times you stopped reading, and you might even sell the misdirection.

Interesting presumption about R1 25-01 being what's talked about, you knowledge cut-off does appear to know R1 update two weeks back was a thing, and that it even improved on function calling.

Of course you have to pretend I meant the former, otherwise "they all have" doesn't entirely make sense. Not that it made total sense before either, but if I say your definition of "they" is laughably narrow, I suspect you will go back to your google contact and confirm that nothing else really exists outside it.

Oh and do a ctrl-f on "irrelevant" please, perhaps some fact grounding is in order. There was an interesting conversation to be had about underpinning of automation somehow without intelligence (Llama 4) but who has time for that if we can have hallucination go hand in hand with forced agendas (free disclaimer to boot) and projection ("doth protest too much")? Truly unforeseeable.

I don't mind the info dump, but I am struggling to connect the relevance of this to topic at hand. I mean, focusing on a single specific capability and generalising it to mean "they all have" caught up with DeepSeek all across the board (which was the original topic) is a reductive and wild take. Especially when it seems to me that this seems more because of misaligned incentive than because it's truly a hard problem.

I am not really invested in this niche topic but I will observe that, yes I agree Llama 4 is really good here. And yet it's a far worse coder, far less intelligent than DeepSeek and that's not even arguable. So no it didn't "catch up" any more than what you could say by pointing out Llama is multimodal but DeepSeek isn't. That's just talking about a different things entirely.

Regardless, I do agree BFCL is not the best measure either, the Tau-bench is more real world relevant. But end of the day, most frontier labs are not incentive aligned to care about this. Meta cares because this is something Zuck personally cares about, Llama models are actually for small businesses solving grunt automation, not for random people coding at home. People like Salesforce care (xLAM), even China had GLM before DeepSeek was a thing. DeepSeek might care so long as it looks good for coding benchmarks, but that's pretty much the extent of it.

And I suspect Google doesn't truly care because in the long run they want to build everything themselves. They already have a CodeAssist product around coding which likely uses fine-tune of their mainline Gemini models to do something even more specific to their plugin.

There is a possibility that at the frontier, models are struggling to be better in a specific and constrained way, without getting worse at other things. It's either this, or even Anthropic has gone rogue because their Aider scores are way down now from before. How does that make sense if they are supposed to be all around better at agentic stuff in tool agnostic way? Then you realise they now have Claude Coder and it just makes way more economic sense to tie yourself to that, be context inefficient to your heart's content so that you can burn tokens instead of being, you know, just generally better.

Not disagreeing with the overarching point but:

That was the easy part

Is a bit hand-wavy in that it doesn't explain why it's only DeepSeek who can do this "easy" thing, but still not Meta, Mistral or anyone else really. There are many other players who have way more compute than DeepSeek (even inside China, not even considering rest of the world), and I can assure you more or less everyone trains on synthetic data/distillation from whatever bigger model they can access.

I tried aidermacs yesterday, it's neat if you use the vterm backend. Does a few things more automatically.