> The company on Wednesday confirmed speculation that the Ox Alpha model is a new iteration of its GLM series and said it will release the weights for it tonight, in response to queries by Bloomberg News.
Seems legit.
It's really hard to know how good it is. So much hype around it.
completely vibes based, but ive been using it to port Mindustry game from Java to C# with agents, and its been working for 50 hours (its 15-20 tks so super slow inference). Its done a fantastic work and its almost finished now. Better results than deepseek flash and gpt luna by a mile on this kind of long term work. Less good than gpt sol or opus. We dont know the param count but my guess is 200-300 range.
likely a distilled glm 5.3 that will punch within 20% of that at 2-3x less size. you'll find that capability is typically very jagged on models that are distilled
I'd be interested to know what was going on with it during the public test as there were numerous reports of it improving considerably at tasks it was asked to do early on in the test compared to later in it.
It's logical to serve the best version (quant) of the model at the beginning so that users keep testing it. It is also reasonable to think that the developer of the model tried to test various quant levels by gradually degrading the model's capabilities.
That benchmark is super sus. Until someone pointed it out, the top performing open weights model was a Kimi K3 fine tune from their sponsor (abacusai/Smaug-Agentic). Now, it's not on the list.
I really want to see hard evidence of distillation before I buy into it. Seems like a lot of sour grapes over not having the sort of lead assumed. In this field, it has been shown repeatedly that leaps in performance come swiftly and without notice.
"Black-Box On-Policy Distillation of Large Language Models", Microsoft Research, https://aka.ms/GAD-project
> 'GAD consistently surpasses standard sequence-level distillation, delivering superior generalization and achieving performance that rivals the proprietary teacher. These results validate GAD as an effective and robust
solution for black-box LLM distillation.'
No RL, although I'm a little bit surprised to see MS Research publishing a paper on distilling GPT5?
That's an interesting paper, but there is virtually no discussion of reasoning behaviors or optimization for long-horizon tasks (i.e., all of the recent advances in LLMs that people care about). The evaluation methodology also is pretty dated:
> We reserve 500 samples of LMSYS-Chat-1M-Clean as the primary test set. We also include test datasets consisting of a 500-sample subset split from Dolly [6], the 252-sample SelfInst dataset [37], and the 80-question Vicuna benchmark [3] to evaluate out-of-distribution generalization. We report the GPT-4o evaluation scores [45, 10], where GPT-4o first generates reference answers and then scores the output of the student model against them. We also conduct human evaluations on the LMSYS-Chat-1M-Clean test set for qualitative assessment.
With the amount of Claudish on the internet now, and in source code repositories (how many Claudish README.mds have you seen?), you don't have to make a single API call to end up with a model that talks like Claude.
And critically, like contracts in general, Anthropic's terms of service is only binding upon the user/counterparty. So even if a company say specifically sought out 'claude-like' content, and claude code traces available on the internet, if they don't use the Anthropic platform there is no ToS claim.
Half of the new open-source stuff on github is written by claude now, all the way from issues to docs. Models are just vacuuming up this dataset during pretraining, naturally picking up the tone. You don't even need direct distillation via api anymore when the whole internet has turned into one big snapshot of Anthropic's weights
Claims about Ox Alpha performing at Fable level were from the social media hype cycle. Everything new in the LLM space brings a wave of influencers hyping it up.
It is a capable small model, but it’s not frontier level. The interesting part will be seeing the model size, how it responds to quantization, and how fast it runs on the kind of non-server hardware that we can buy without selling a kidney.
They don't have to be getting paid. The natural bias of media is towards laziness and sensationalism (stolen from Jon Stewart, so maybe the same is true about comment sections).
Kinda useless to compare simply based on model without considering harness. Different agents handle the context etc completely differently. I would like to start seeing these model vs model comparisons across different harnesses.
omp+0x-alpha beat both cc+fable and codex-sol in creating/refactoring a big eval setup. the former just knows where things should belong and completed the task all the way while the other two failed on both metrics.
I've been having oxa and sol do architecture design then compare notes. Sol is definitely still way ahead. But there's reliably some really good wins ideas and concepts that OxA throws out there that Sol is very happy to encorporate.
One thing that I think matters a lot for the non developers, all three of us (sol, oxa, and me) usually agree that oxa's write up is far far better. It explains the situation very well, and has great structure for its write ups. Sol gets the job done, but it's terrible at re-explaining the problem for humans, at laying out information. It also doesn't show it's thinking, so it's imo a terrible peer to work with!
Xi has made it official policy, see his keynote speech at their World AI Conference last month:
> We should seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing. [1]
People have pointed that this seemingly made Alibaba/Qwen turn around from closing their models (this was rumored after the shakeup early this year [2]) and release the weights for even the Max variant of their new models, which they previously did not.
Well, China is at their usual barely legal (at least if the WTO would be worth even a bit after it got thoroughly gutted) game, just burning their effectively infinite cash reserves to undercut Western providers and eventually force them out of business.
In the end there's barely any moat that any of the ludicrously "valued" AI companies have - the only thing justifying the valuations of SpaceX/xAI/Tesla/Anthropic/OpenAI is a supposed "secret sauce" that, frankly, barely exists any more.
Everyone and their dog can go and run LLMs for dirt cheap on their own hardware.
That is saying a lot if Ox Alpha is also small and relatively cheap computationally. I hope so; I love deepseek-v4-flash-0731 and use it frequently. Fast inference is good and fits with my dev style: I like to be in the loop, not let an agent code on its own for long periods of time.
From their blog post, it's 320B total parameters and 18B active parameters, so a similar size, but slightly bigger.
Regular pricing is $0.15 input, $0.50 output... but currently 50% off, making it $0.075 input and $0.25 output. That beats most of the V4 Flash providers, but not all, and obviously tokens per task may not be equivalent.
I've also just noticed the blog post reveals the Artificial Analysis score - it's a 57, so it's Opus 4.8 / 5.6 Terra level.
Z.AI is the only provider for GLM 5.3 on OpenRouter. I don't see 5.3 on Hugging Face. Not sure if this new model is "full GLM" or something smaller, or if they will like Moonshot AI publish weights but put restrictive license [1], which will again leave Z.AI as single GLM model provider on OpenRouter.
I am happy if they publish under restrictive license. Developing model costs a tons of money and company need to make money somehow by still open sourcing project.
There's a lot of brand confusion among the Chinese models right now. Kimi, Qwen, GLM, Z.ai, Ox. We might know the difference (or I should say, someone does because I'm losing track already) but these models have no chance at end user penetration and loyalty until there's a single focused survivor.
It took me a year talking about it until my wife knew that ChatGPT and Gemini are two different things.
I have seen studies from MIT and Stanford that the majority or US startups are using much less expensive open weight models so consumers of their products are open model users whether they know it or not. These are often Chinese models.
Not to go off topic but I am pleased to see open model support from US companies like Poolside.ai, NVIDIA, IBM, Google, etc.
There's a lot of brand confusion among the American models right now. ChatGPT, Claude, Gemma, OpenAI, Meta, Google, Muse Spark, Anthropic, Microsoft, Gemini. We might know the difference (or I should say, someone does because I'm losing track already) but these models have no chance at end user penetration and loyalty until there's a single focused survivor.
It took me a year talking about it until my wife knew that Kimi K3 and GLM 5.3 are two different things.
The bubbling froth at the open edge is getting user adopted at a crazy pace, by the early adopter persona trying them all within hours to days. This persona loves taking apart and putting together novel things, and telling others.
Fast follower persona clusters around emerging zeitgeist across the tellings. At the moment, arguably that's mostly Qwen for everyday hobbyists, and GLM for those that can run 512GB to 1.5TB of memory. This persona is seeking viable applied results: "I have frontier at home".
The early majority pick things up after models are curated into apps like LM Studio or one's platform app of choice, usually at least one major release behind because it takes that long to choose and package into mass distribution.
This is the step where early majority persona "has no idea" what the parade of weird names is about, they care about qualia of the conversations they try to have.
This persona is, at present, very under-served, and likely to remain so until mass devices can perform feeling like 27B at Q4 large quality better, or workplace devices can achieve a pragmatic utility like 135B at Q8 or better.
Harnesses that work where the workplace persona lives bridge this. This persona doesn't care the Chinese model name, they care "does it code?" For that, the applied harness and model take time to be matched, as JetBrains did harnessing a tailored Qwen 3.6 in the IDE. More efforts like https://www.jetbrains.com/junie/ are needed for the majority persona to perceive value from changing their workflow again.
HN's "job" is better outcomes with less friction at each persona.
In raw numbers of humans... "the early majority" surely would be those that use ChatGPT or Gemini (aka Google) and pay between $0 and $20 a month?
I would be surprised if the specialist that knows that various Chinese models exist and/or that a user might choose a harness and model separately are a "majority" even of the early variety... in terms of revenue, humans, tokens, or any metric.
> these models have no chance at end user penetration and loyalty until there's a single focused survivor.
This reminds me a lot of media horse-race reporting, saying that "candidate X has no chance unless they" and "candidate Y has a strong showing in", and it's very thinly cover for the publication liking Y and disliking X, avoiding talking about actual policy, and trying as much as they can to make their predictions self-fulfilling.
> have no chance at end user penetration and loyalty until there's a single focused survivor.
But why does that matter? End users (I believe, feel free to correct) do not really contribute all that much revenue-wise.
They're certainly not the SOTA target audience.
The professional market doesn't need a household name. They need the most sensible tool for the job, and the CN models right now tick many boxes when it comes to that.
There will never be any significant loyalty in this field, models are commodities, we will keep switching to the best/most convenient for a long time.
Proper usage in applications will go through similar considerations.
The only "captive" users will be non-tech enterprise users, but they are already in Gemini/Copilot land because they are natural extension to existing Google Cloud/Teams plans and nobody cares about what models are inside those, procurement, contracts and data retention are what matters.
Only reason people are interested is it’s free at the moment. I wasn’t impressed by its performance. Once the model gets a price tag it’s usage will be negligible.
Look, it's kind of like that 1982 HydraTech 16ft bass boat with a pearl glitter paint job thats on Craigslist. When the wife asks, you low balled it and they accepted - an offer too good to refuse.
I have many rigs, but let's take 1 for example. 160gb. $1000 that's what it cost. 10 16gb MI50 gpus from ebay at $90. $900. Plugged them into an $100 octominer case from Facebook marketplace. I'm sure that doesn't satisfy you, keep coming up with excuses instead of finding ways to make this happen for you. You either find a way to get in and play or you sit on the sideline and moan about those in the field.
> The company on Wednesday confirmed speculation that the Ox Alpha model is a new iteration of its GLM series and said it will release the weights for it tonight, in response to queries by Bloomberg News.
Apparently someone working at a 3rd party inference provider also got confused and posted confirmation about it being a glm-flash model, despite having an embargo on that info. Someone jumped in the comments and told them they missed the timezone :)
In any case it should be releasing in a few hours. Timezones are hard.
Rather than a pelican, for fun I showed it a couple of screenshots from Niu Lai and asked it to create an SVG inspired by the images. I explained a little about how the movie had been made by a mother & son team, initially derided but then went on to surprise cult box office success. It came up with this:
Created in a single turn - but technically not a "one-shot", because I gave it a tool to convert SVG to PNG so it could visualize what it had made. I asked it to keep iterating with tools during the same turn until it was happy.
I've also been using Ox Alpha for tasks that better resemble real work, and I'm really enjoying working with it. I've downgraded my Anthropic account so I can put some budget towards Ox Alpha instead, with the rumors that this one is going to be cheap. Opus & Fable are still better at getting large tasks / features done autonomously, but Ox Alpha can work autonomously too, and it's fun. I'm enjoying working with Ox in a way that I'm just not enjoying talking to the 5.0 Anthropic models. (As much as I don't want to say that, as someone with Claude /stickers on their laptop.)
Yeah, that's a fair point. "More intelligible than Adriano Celentano in Prisencolinensinainciusol" is not a high bar.
As another comparison, I went back to MiniMax M3 for a while last night. It was significantly faster than Ox, but I felt M3's replies were harder to parse, not quite getting to the point. But I guess I could curb that with some prompts.
It depends if the Ox Alpha pricing is as cheap as was being rumored. If it's competitive with DeepSeek Flash and significantly undercutting Luna, that feels like it will be significant.
I had Ox Alpha working on coding tasks for a couple days non-stop, via OpenRouter and OpenCode Zen. It was able to complete tasks at a level that I'd put between Sonnet and Opus. It makes few mistakes, but is not that smart.
The main issue for me, is that it degraded into a doom loop several times. One of them was running the same bash command about a thousand times. The last model I've used that had this problem was Mimo 2.5, which is quite dated at this point. As a result of this, you cannot leave it unattended / not usable for agents.
I couldn't get past all the network errors on OpenCode. Seemed smart enough, and was useful when I was low on usage on Claude, but beyond that, really hard for me to say whether it was Good or Bad.
I asked it to create a design system in Paper, and it actually did a fairly decent job when it wasn't getting network errors. I'd say it's much, much closer to good than bad.
Doom loops are as much of a model problem as it is deficiency of the harness. I have not seen any other open source harness that deals with them except the one I started because of this obvious gap.
See my other comment with examples where 0x Alpha is working non-stop on various projects with zero problems.
It feels like all these models are missing a long horizon orchestrator of some kind. Humans don't think in a single thread, they have thoughts that appear, pause and disappear.
The reason humans don't go into a doom loop is because of this ability of other thoughts to interject and say, no this is stupid.
I'm not 100% sure what I'd do to detect this for shells specifically, maybe prompt the user, but my toy harness scolds refuses to re-read files because the previous iteration of GLM was room-looping on that. Clearly something that they need to work on.
Pi.dev has an extension that does it. None of the models I use seem to have that failure mode so I haven't bothered throwing it in.
The failure mode I run into commonly is agents just stop sometimes. Even sending a "." Or something they start back up, but I haven't worked out exactly how to fix that generally in harness, bit unclear how to tell if they're done or just derped to a stop.
I know that AI is moving fast, but Mimo 2.5 literally came out four months ago. I literally had to double-check after I read this because it felt like just yesterday.
Also, I'm not sure what the industry standard is right now, but my (self-built) harness automatically exits with an error code whenever it detects similar tool calls being sent. It's pretty easy to set that kind of thing up.
I’m more curious on the size. If it’s smaller than or equal size to GLM 5.3, this would be a crazy good model. If it’s closer to deepseek pro, it would be a good model. If it’s near Kimi K3, I think it’s ok.
Definitely agree. If it is small (eg. Qwen 3.8 28b or gpt-oss-120) then this might be amazing. If it is anywhere near Kimi K3 it would need to have some other differentiating factor than intelligence.
Calling it now: The big deal about this model is the sheer volume they were offering through openrouter and OpenCode. How? Chinese AI accelerators / nvidia-free stack
Ox alpha is better at UI than GPT 5.6 Sol. Not a high bar considering Sol sucks at UI, but as someone who just has a codex sub, I've used almost 1B tokens of ox alpha these last few days to complement Sol smartness.
Inference was atrocious in terms of speed and constant timeouts. If it's served fast it will be a delight to use.
128 comments
[ 0.76 ms ] story [ 74.0 ms ] threadSeems legit.
It's really hard to know how good it is. So much hype around it.
Wenghi is behind deepswe, one of the best benchmarks. 63% at deepswe.
https://livebench.ai/
while here it outperforms Fable by a significant margin:
https://oxalpha.com/
but if the latter is true, will people still say it was "distilled" from Fable?
Source: https://twitterwebviewer.com/?tweet=2091116504787935350
> 'GAD consistently surpasses standard sequence-level distillation, delivering superior generalization and achieving performance that rivals the proprietary teacher. These results validate GAD as an effective and robust solution for black-box LLM distillation.'
No RL, although I'm a little bit surprised to see MS Research publishing a paper on distilling GPT5?
> We reserve 500 samples of LMSYS-Chat-1M-Clean as the primary test set. We also include test datasets consisting of a 500-sample subset split from Dolly [6], the 252-sample SelfInst dataset [37], and the 80-question Vicuna benchmark [3] to evaluate out-of-distribution generalization. We report the GPT-4o evaluation scores [45, 10], where GPT-4o first generates reference answers and then scores the output of the student model against them. We also conduct human evaluations on the LMSYS-Chat-1M-Clean test set for qualitative assessment.
The metric used there is me screaming at my screen per operating hours.
Does it matter? IMO not really. Weights are open after all. (Or.. soon at least for 5.3)
And critically, like contracts in general, Anthropic's terms of service is only binding upon the user/counterparty. So even if a company say specifically sought out 'claude-like' content, and claude code traces available on the internet, if they don't use the Anthropic platform there is no ToS claim.
It is a capable small model, but it’s not frontier level. The interesting part will be seeing the model size, how it responds to quantization, and how fast it runs on the kind of non-server hardware that we can buy without selling a kidney.
They claim an "independent community benchmark" (pass-fail evaluation on tasks) here: https://oxalpha.com/ox-alpha-vs-fable-5
One thing that I think matters a lot for the non developers, all three of us (sol, oxa, and me) usually agree that oxa's write up is far far better. It explains the situation very well, and has great structure for its write ups. Sol gets the job done, but it's terrible at re-explaining the problem for humans, at laying out information. It also doesn't show it's thinking, so it's imo a terrible peer to work with!
on toy benches it made quite a few mistakes but was able to fix all of them on its own
(meaning more tokens, more turns, more tool calls — but same outcome as gpt 5.6 sol)
Didnt know they exist - looks very good, maybe even better than Archive.ph
Chinese labs are not releasing all of their model weights. Qwen is known as an open weight model by most, but their top model is not open weight.
Releasing weights is a marketing strategy for newer labs to get their brand out there.
https://mashable.com/article/chatgpt-gpt-4o-ai-retirement-pr...
> We should seize this rare, historic opportunity to encourage open source, openness, collaboration and sharing. [1]
People have pointed that this seemingly made Alibaba/Qwen turn around from closing their models (this was rumored after the shakeup early this year [2]) and release the weights for even the Max variant of their new models, which they previously did not.
1: http://english.scio.gov.cn/topnews/2026-07/18/content_118605...
2: https://simonwillison.net/2026/Mar/4/qwen/
In the end there's barely any moat that any of the ludicrously "valued" AI companies have - the only thing justifying the valuations of SpaceX/xAI/Tesla/Anthropic/OpenAI is a supposed "secret sauce" that, frankly, barely exists any more.
Everyone and their dog can go and run LLMs for dirt cheap on their own hardware.
Regular pricing is $0.15 input, $0.50 output... but currently 50% off, making it $0.075 input and $0.25 output. That beats most of the V4 Flash providers, but not all, and obviously tokens per task may not be equivalent.
I've also just noticed the blog post reveals the Artificial Analysis score - it's a 57, so it's Opus 4.8 / 5.6 Terra level.
https://z.ai/blog/glm-5.3-flash
Z.AI is the only provider for GLM 5.3 on OpenRouter. I don't see 5.3 on Hugging Face. Not sure if this new model is "full GLM" or something smaller, or if they will like Moonshot AI publish weights but put restrictive license [1], which will again leave Z.AI as single GLM model provider on OpenRouter.
[1] https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE
It took me a year talking about it until my wife knew that ChatGPT and Gemini are two different things.
Not to go off topic but I am pleased to see open model support from US companies like Poolside.ai, NVIDIA, IBM, Google, etc.
It took me a year talking about it until my wife knew that Kimi K3 and GLM 5.3 are two different things.
Fast follower persona clusters around emerging zeitgeist across the tellings. At the moment, arguably that's mostly Qwen for everyday hobbyists, and GLM for those that can run 512GB to 1.5TB of memory. This persona is seeking viable applied results: "I have frontier at home".
The early majority pick things up after models are curated into apps like LM Studio or one's platform app of choice, usually at least one major release behind because it takes that long to choose and package into mass distribution.
This is the step where early majority persona "has no idea" what the parade of weird names is about, they care about qualia of the conversations they try to have.
This persona is, at present, very under-served, and likely to remain so until mass devices can perform feeling like 27B at Q4 large quality better, or workplace devices can achieve a pragmatic utility like 135B at Q8 or better.
Harnesses that work where the workplace persona lives bridge this. This persona doesn't care the Chinese model name, they care "does it code?" For that, the applied harness and model take time to be matched, as JetBrains did harnessing a tailored Qwen 3.6 in the IDE. More efforts like https://www.jetbrains.com/junie/ are needed for the majority persona to perceive value from changing their workflow again.
HN's "job" is better outcomes with less friction at each persona.
I would be surprised if the specialist that knows that various Chinese models exist and/or that a user might choose a harness and model separately are a "majority" even of the early variety... in terms of revenue, humans, tokens, or any metric.
(Happy to be proven wrong)
This reminds me a lot of media horse-race reporting, saying that "candidate X has no chance unless they" and "candidate Y has a strong showing in", and it's very thinly cover for the publication liking Y and disliking X, avoiding talking about actual policy, and trying as much as they can to make their predictions self-fulfilling.
But why does that matter? End users (I believe, feel free to correct) do not really contribute all that much revenue-wise. They're certainly not the SOTA target audience.
The professional market doesn't need a household name. They need the most sensible tool for the job, and the CN models right now tick many boxes when it comes to that.
Proper usage in applications will go through similar considerations.
The only "captive" users will be non-tech enterprise users, but they are already in Gemini/Copilot land because they are natural extension to existing Google Cloud/Teams plans and nobody cares about what models are inside those, procurement, contracts and data retention are what matters.
Related PR: https://github.com/jeffhajewski/latticedb/pull/5
The session used ~100K input tokens, ~60K output tokens, and ~80K thinking tokens.
I reviewed it using gpt-sol-medium, and it seems to be satisfied with it's work.
Where? And "Tonight" in which timezone?
Apparently someone working at a 3rd party inference provider also got confused and posted confirmation about it being a glm-flash model, despite having an embargo on that info. Someone jumped in the comments and told them they missed the timezone :)
In any case it should be releasing in a few hours. Timezones are hard.
https://x.com/syneryder/status/2091978367579156569/photo/1
Created in a single turn - but technically not a "one-shot", because I gave it a tool to convert SVG to PNG so it could visualize what it had made. I asked it to keep iterating with tools during the same turn until it was happy.
I've also been using Ox Alpha for tasks that better resemble real work, and I'm really enjoying working with it. I've downgraded my Anthropic account so I can put some budget towards Ox Alpha instead, with the rumors that this one is going to be cheap. Opus & Fable are still better at getting large tasks / features done autonomously, but Ox Alpha can work autonomously too, and it's fun. I'm enjoying working with Ox in a way that I'm just not enjoying talking to the 5.0 Anthropic models. (As much as I don't want to say that, as someone with Claude /stickers on their laptop.)
hard agree. it does not really feel "smart", but the personality is super refreshing
That’s very valid, but right now every other model I use is easier to talk to than Opus 5.0
Opus 5.0 has an impenetrable way of communicating. I can parse it, but it takes so much more work than it should.
As another comparison, I went back to MiniMax M3 for a while last night. It was significantly faster than Ox, but I felt M3's replies were harder to parse, not quite getting to the point. But I guess I could curb that with some prompts.
It depends if the Ox Alpha pricing is as cheap as was being rumored. If it's competitive with DeepSeek Flash and significantly undercutting Luna, that feels like it will be significant.
The main issue for me, is that it degraded into a doom loop several times. One of them was running the same bash command about a thousand times. The last model I've used that had this problem was Mimo 2.5, which is quite dated at this point. As a result of this, you cannot leave it unattended / not usable for agents.
An amusing thought of returning to your workstation to find it as an obsidian block after it gets stuck executing "dd" thousand times.
It sounds like you were using a quant model.
See my other comment with examples where 0x Alpha is working non-stop on various projects with zero problems.
The reason humans don't go into a doom loop is because of this ability of other thoughts to interject and say, no this is stupid.
The failure mode I run into commonly is agents just stop sometimes. Even sending a "." Or something they start back up, but I haven't worked out exactly how to fix that generally in harness, bit unclear how to tell if they're done or just derped to a stop.
I know that AI is moving fast, but Mimo 2.5 literally came out four months ago. I literally had to double-check after I read this because it felt like just yesterday.
Also, I'm not sure what the industry standard is right now, but my (self-built) harness automatically exits with an error code whenever it detects similar tool calls being sent. It's pretty easy to set that kind of thing up.
[0]: https://openrating.io/blog/current-state-of-ai-model-fingerp...
Inference was atrocious in terms of speed and constant timeouts. If it's served fast it will be a delight to use.