Blog / Model guides

Is Ox Alpha open weights or open source? Yes, since 26 August

Published 25 August 2026 · Updated 27 August 2026

Yes. Ox Alpha is GLM-5.3-Flash, a 320B-parameter mixture of experts from Z.ai, and its weights have been on Hugging Face as zai-org/GLM-5.3-Flash under the MIT license since 26 August 2026. Z.ai's documentation states: "Before release, we tested GLM-5.3-Flash anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback." The company told Bloomberg the same day that the weights would ship that evening, and the repository was published that night. The checkpoint and the license are public; the training data and code are not, so the model is open weights rather than open source.

What Ox Alpha was

Ox Alpha appeared on OpenRouter on 20 August 2026 as stealth/ox-alpha, free, with no vendor attached. OpenRouter described it as a reasoning model for coding and sustained agentic work, "developed and operated by a third-party provider who has chosen to remain anonymous during this preview". The same listing now reads "developed and operated by ZAI, revealed to be ZAI GLM-5.3-Flash" and points at the production entry, z-ai/glm-5.3-flash, created 26 August.

Bloomberg reports that the anonymous model topped OpenRouter's usage leaderboard and more than doubled DeepSeek's traffic there during the week it ran without a name.

How the model was identified before the announcement

Before Z.ai's announcement, kingy.ai published an analysis attributing the model to the GLM family. Token counts from the endpoint matched GLM-5.3 plus a fixed 75-token offset, across English, German, Chinese, code and emoji. Turning reasoning off produced the same error message GLM-5.3 returns.

The same analysis listed two unresolved points: Ox Alpha accepted image and video input, which the public GLM-5.3 did not, and no product called GLM-5.3 Flash appeared on any Z.ai page. The release explains both. GLM-5.3-Flash is a separate, natively multimodal model, and Z.ai had not announced it. TechCrunch had noted a competing reading that pointed at an unreleased Microsoft MAI model, and earlier guesses named Gemini. Neither was correct.

Which Hugging Face repositories are official

Z.ai publishes two: zai-org/GLM-5.3-Flash, the FP8 checkpoint, and zai-org/GLM-5.3-Flash-BF16. The repositories named after Ox Alpha that appeared during the preview are unchanged. The most visible, Ox-Alpha-GGUF, states in its card that it holds no weights, and the others do not name a source. A conversion made from the Z.ai checkpoint would state which GLM-5.3-Flash revision it was made from.

What it takes to run

The native FP8 checkpoint is about 306 GiB of weights. Z.ai's vLLM recipe calls for Hopper or newer GPUs, tensor parallelism of at least 4 and vLLM 0.29; derived from that, 4x H200 or 8x H100 80 GB fits with room for KV cache, and the BF16 variant needs an 8x H200 node. It does not fit a single GPU. The full breakdown, including Z.ai's numbers against DeepSeek V4 Flash, is in GLM-5.3-Flash hardware requirements and license. GLM-5.3-Flash is a separate model from the GLM-5.3 flagship, whose weights are listed for 28 August; see the GLM-5.3 page.

The cost of that node, estimated from published rates: the H200 rate is the $7 to $14 per hour observed for 2x H200 on RunPod in August, scaled to four GPUs; the AWS figure is the p5en.48xlarge list price. Months are 730 hours, business hours 176.

Data handling during the preview

The preview listing stated that prompts and completions were retained by the provider and not used for training. That provider is now identified as Z.ai. The models that run in your own cloud account are listed in the catalog.

Questions people ask

Start a 7-day free trial