Behaviorally fingerprinting Ox Alpha's provenance
ctgt.aiError messages matching Z.ai GLM I think are the simplest/most compelling. I had Opus 4.8 poke it and it came back with a couple different errors than the article mentions.
Matching the tokenizer is interesting tho
I find the tokenizers most compelling. That's what the model is trained on, it's an immutable fact of the model and its architecture. You know for a fact that the model is at least related to other models that way. And if a tokenizer is unique / specific to one lab, like GLM's is, it's basically as good as it gets.
Comparatively, you can't be 100% sure that Z.ai isn't able to host some other lab's model (although in this case, the hosting errors still support the GLM theory).
No, that's not how anything works.
You can finetune an LLM to a new tokenizer by nudging just a few layers (even wildly different kinds of tokens), and there's nothing stopping a lab from using someone else's tokenizer.
Recent and related:
Ox-Alpha Is GLM? - https://news.ycombinator.com/item?id=49422226 - Aug 2026 (65 comments)
A mysterious free AI model is impressing developers. Nobody knows who made it - https://news.ycombinator.com/item?id=49406289 - Aug 2026 (4 comments)
Ox Alpha - https://news.ycombinator.com/item?id=49381896 - Aug 2026 (202 comments)
I was using it side by side with GLM 5.3, and they were very, very similar. Also, the new zhipu 1 GW data center plus a flash(smaller?) model, can justify the 100T/day they said they were able to serve.
Pretty cool model, especially since it's not a nanny, if you want to unlock your own devices, like rooting an Android, it will happily help instead of flagging you.
Available for free via OpenRouter and Nous free tier. Also via OpenCode Go, but you have to pay a $5–$10 subscription. APIs are hammered now, so service is bumpy.
Another strong hint is that the uptime graph of GLM-5.3 by Z.ai is very similar to that of Ox Alpha:
https://openrouter.ai/stealth/ox-alpha#uptime
https://openrouter.ai/z-ai/glm-5.3#uptime
Screenshot of a recent blip: https://files.catbox.moe/haq90y.png
Can someone explain why people care about that?
Both as in "Why is there a stealth launch like that in the first place?" but also "Why does it matter? Is it very good in something?"
Stealth launch: builds hype, allows them to collect user preference data and see where the model fails. Why people care: it's free, decent, and people love a good mystery.
Aaah, free inference. Yeah that checks out and fits very well with weird Internet hype.
Okay, fair enough. Thanks!
It's a good model ser.
I think the blind test is valuable. To what degree is our impression of a model affected by who created it?
I agree, the blind test is valuable - like the old Pepsi Challenge.
I like Ox Alpha. I'm getting some actual work done with it, I'm really enjoying working with it, and I've already cut back my Anthropic budget in anticipation of working with this model instead.
But I'm terrified that when the model is revealed, I'll discover that it was Grok 4.7 all along.
I think there was some loud speculation is was Google. It's cool people can pretty definitively show it's not. But also maybe Google is being crazy sneaky and faking it! (Joking)
I think the stranger thing is that people spend tens of hours doing analysis like this to hit an inevitably-expiring hype cycle that will give us a definitive answer shortly.
Don't you think the forensics is fun?! This is exactly the sort of thing i would've expected hn to be broadly interesting to hn. It's a complicated technology and people are poking it and learning things
It's fun! Also it's an interesting commentary on where the AI industry is as a whole.
It's decent. How good depends on compute needs.
Speculation that this is GLM 5.3 Flash.
Very likely, NYT confirmed it's releasing Friday.