Blog / Model guides
Is Ox Alpha open weights or open source? What we know
No. As of 25 August 2026, Ox Alpha is not open weights and it is not
open source. Nobody has published a checkpoint, a license or any code,
and nobody has said who built it. It exists only as a hosted endpoint,
stealth/ox-alpha on
OpenRouter, free
during a preview, and the party serving it has chosen not to identify
itself. The rest of this post is what can be established around that
fact, with sources.
What Ox Alpha is
Ox Alpha appeared on OpenRouter on 20 August 2026 with no vendor attached. OpenRouter describes it as a reasoning model for coding, sustained agentic work and production workloads, and says it is "developed and operated by a third-party provider who has chosen to remain anonymous during this preview". OpenRouter routes the requests; it is not the developer. Everything below is from the listing.
| What | Listed value |
|---|---|
| First seen | 20 August 2026 |
| Price | Free during the preview |
| Context window | 1,048,576 tokens |
| Completion cap | 131,072 tokens |
| Input | Text, images and video; text out |
| Features | Tool calling, structured JSON output, reasoning that cannot be switched off |
| Data handling | Prompts and completions are retained by the anonymous provider, not used for training |
The quality of the model against that price of zero is why it took over developer discussion within days; TechCrunch covered the guessing game three days after launch.
Why people think it is GLM-5.3
The leading attribution comes from community fingerprinting, laid out in detail by kingy.ai. Two observations carry most of the weight. Token counts from the endpoint track GLM-5.3 exactly, plus a fixed 75-token offset, across English, German, Chinese, code and emoji, and tokenizers are distinctive in how they split unusual input. And when you try to turn reasoning off, Ox Alpha returns the identical error message GLM-5.3 returns: reasoning is always on, and the accepted effort levels are low, high and max. Matching prose style is easy to dismiss; matching error contracts is not.
The gaps matter too, and the same analysis lists them. Z.ai uses one tokenizer across GLM-5.2 and GLM-5.3, so token counts cannot separate the two. Ox Alpha accepts images and video, which the public GLM-5.3 materials do not emphasise. And no product called GLM-5.3 Flash exists on any Z.ai page. There is also a competing theory: TechCrunch notes a reading of the evidence that points at an unreleased Microsoft MAI model instead, and the earliest guesses said Gemini. Z.ai has confirmed nothing, and neither has anyone else.
The Ox Alpha files on Hugging Face are not weights
Searching Hugging Face for Ox Alpha already returns results, which is how "is Ox Alpha open weights" starts trending in the first place. Look closely and there is nothing in them. The most visible repository, Ox-Alpha-GGUF, says in its own card that it is a community placeholder with no weights uploaded. Others carry no provenance at all. The logic is simple: a model that exists only behind an API cannot be converted into a GGUF or safetensors file without an authorized checkpoint, so until a named developer publishes one under a stated license, any file claiming to be Ox Alpha deserves suspicion.
When could open weights appear
If the GLM-5.3 attribution is right, there is a dated clue. Z.ai launched GLM-5.3 on its own API on 14 August 2026 and said the open weights follow about two weeks later, after a safety evaluation, which points at late August. GLM-5.2's weights shipped under MIT; the GLM-5.3 license is unconfirmed. Our GLM-5.3 page tracks that release and the hardware it will need. Note the limit of the inference: a GLM-5.3 checkpoint on Hugging Face would not by itself tell you what Ox Alpha is. Stealth previews usually end with the developer claiming the model, and Z.ai has run one before, so the cleanest confirmation is simply waiting for a name.
If the Microsoft MAI theory is right instead, there is no stated open-weights plan at all. And if it is neither, there is nothing to go on.
If you need something you can self-host today
Ox Alpha cannot run in your own cloud account, and the free preview has a cost worth naming: the listing states that your prompts and completions are retained by a provider whose identity you do not know. That is fine for experiments, but think twice before sending it anything sensitive. The models that can run privately today, with measured boot times where we have them, are in the catalog; DeepSeek V4 Flash is our most popular model.
Questions people ask
Is Ox Alpha open source?
No. As of 25 August 2026 there are no published weights, no license, no code and no named developer. It is only available as a free hosted preview on OpenRouter.
Are the Ox Alpha files on Hugging Face real weights?
No. The repositories named after it are placeholders or uploads with no provenance; the most visible one states outright that it holds no weights. An API-only model cannot become a legitimate GGUF without an authorized checkpoint.
Who makes Ox Alpha?
Nobody has said. Token counts and an identical reasoning error message point at Z.ai's GLM family; a competing reading points at an unreleased Microsoft MAI model. No company has claimed it.
Can I run Ox Alpha on my own GPUs?
Not today; it is API-only. If the GLM-5.3 attribution proves correct and Z.ai publishes the GLM-5.3 weights as stated, a closely related model would become self-hostable. That is what the evidence suggests, not something anyone has announced.