Z.AI confirms Ox Alpha is a GLM-series model and plans to release its weights

I couldn't find independent sources at first, but Bloomberg says Z.AI officially confirmed it. The claims about performance still have a lot of hype around them.

I'm trying to find out the model size. Nobody seems to know the parameter count yet.

I've run it on a Java-to-C# Mindustry port with agents for about 50 hours. It is slow at 15 to 20 tokens per second, but it has done much better than DeepSeek Flash and GPT Luna on long-running work, though worse than GPT Sol or Opus. My guess is 200 to 300B parameters.

The only concrete capability figure I've seen is 63% on DeepSWE, cited by Wenghi. I'd rather see more reliable evaluations than social posts.

Reports that it got worse later in the public test may just be pass@K variance. I find Chinese open-weight models vary more between runs than heavily RL-trained models like Fable or Opus.

They could also have changed the serving quantization during the test. A provider has an incentive to begin with the best quantized version, then try lower-quality quants.

LiveBench puts it below GPT-5.4 Nano, while oxalpha.com claims it beats Fable by a wide margin. Those results conflict too much to draw conclusions.

That oxalpha.com benchmark is not official and its apparent Fable result was an unfinished run. I would not treat it as evidence.

I want hard evidence that this was distilled from Fable before accepting that narrative. Performance jumps happen quickly in this field, and the distillation claim may be sour grapes.

The Fable-level claims came out of the social-media hype cycle. I expect a capable small model, not a frontier model. Size, quantization behavior, and local inference speed matter more.

A model-only comparison is mostly useless without the agent harness. Context handling differs enough that the same model can look very different across setups.

I saw omp plus Ox Alpha beat cc plus Fable and Codex Sol on building and refactoring a large evaluation setup. It completed the task, while the others failed both metrics.

Chinese labs do not release every model's weights. Open weights are a marketing tactic for labs that want recognition, and even Qwen's top model is not open weight.

Xi called for open source, openness, collaboration, and sharing at the World AI Conference. That policy may have pushed Alibaba and Qwen toward releasing weights for newer Max models.

Z.AI releasing weights sounds good, but GLM 5.3 is still only available from Z.AI on OpenRouter. The license and whether Ox Alpha is a full GLM model or a smaller one will matter.