Ox Alpha Was GLM-5.3-Flash: Z.ai Ships MIT Weights After a 62-Trillion-Token Stealth Run

Z.ai claimed Ox Alpha on August 26. It is GLM-5.3-Flash, a 320B-A18B natively multimodal mixture-of-experts model, the weights are MIT on Hugging Face, and Z.ai says the entire anonymous preview ran on Chinese AI chips. The slug stealth/ox-alpha is already gone from OpenRouter's catalog.

What Z.ai actually confirmed

• Open weights, MIT. config.json reports 45 layers, 288 routed experts with 8 active per token, a 154,880 vocabulary, FP8 weights, and max_position_embeddings of 1,048,576. Video and image token IDs are in the config, so the multimodality is in the weights, not just the API.
• Served on domestic silicon. Z.ai's words are "a large-scale cluster of Chinese AI chips" and "tens of thousands of domestically developed accelerators." It does not name a vendor. The 100,000-chip figure in circulation is the South China Morning Post's characterization, not Z.ai's.
• A custom stack, built partly by the model. Z.ai wrote a dedicated inference engine on top of SGLang, with Encode-Prefill-Decode disaggregation, W8A8, and hybrid INT8/FP8/BF16 cache quantization. It reports a 3x end-to-end serving gain over its own baseline on the same hardware, and says its GLM-5.3-powered infrastructure agent helped write and tune the kernels.

About that 100 trillion tokens per day

That number was never a measurement. It was OpenCode's gateway capacity claim at launch, and most coverage has been repeating it as throughput. Here is the difference, using OpenCode's own public dashboard.

For scale: OpenCode's entire platform, every model from every lab combined, has never cleared about 22 trillion tokens in a single day. The advertised ceiling was roughly five times the busiest day the platform has ever had. The real result is still the headline, though. GLM-5.3-Flash sits at number one with 45T tokens, two and a half times the 18T of second-place DeepSeek-V4-Flash.

Why it matters if you build on open models

Two things. First, MIT weights on a 320B-A18B multimodal MoE with a real 1M context, at $0.15 per million input tokens and $0.50 per million output on the list price, with several OpenRouter providers running a 50 percent launch discount. Twelve providers were serving it within a day, which is what open weights buy you and what a stealth endpoint never can.

Second, the hardware claim. Z.ai says it reached per-token cost comparable to mainstream NVIDIA GPUs on domestic accelerators. Treat that as vendor-reported, because it is, and none of us can audit it. But the workload behind it was not a lab benchmark. It was a week of real developer traffic, and that is a harder thing to fake than a chart.

Our scorecard

Four days ago we published a fingerprint of the unclaimed model and declined to name the lab. The serving-profile match against the GLM line held. The competing tokenizer analysis pointing at Microsoft did not. And the benchmark we corrected, the retracted 80 percent DeepSWE score that we re-reported at roughly 63 percent, matches the 63.4 on DeepSWE v1.1 that Z.ai has now published itself. Z.ai also reports 48.8 on AutomationBench against GLM-5.2's 26.2, and 29.0 against Claude Opus 4.8's 29.5 on its in-house Code Bench. All vendor-reported. We have not tested this model.

The base rate from our Hunter Alpha coverage still stands: every tracked cloaked model eventually gets claimed, and roughly two in eight first community guesses are wrong.

Run it today

# Weights (MIT)
huggingface-cli download zai-org/GLM-5.3-Flash

# Or hit it through OpenRouter, no weights needed
curl https://openrouter.ai/api/v1/chat/completions \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"z-ai/glm-5.3-flash","messages":[{"role":"user","content":"hello"}]}'

# Check the catalog entry yourself
curl -s https://openrouter.ai/api/v1/models \
  | jq '.data[] | select(.id=="z-ai/glm-5.3-flash")'

Serving recipes exist for SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth. At 320B in FP8 this is not a laptop model, so if you only want to try the behavior, the API path above costs a few cents. One thing worth doing in the next ten minutes: if you sent Ox Alpha anything sensitive during the free preview, go re-read the Stealth Program terms you accepted. The perpetual training license in them did not expire when the model got a name.

Sources

• Z.ai, GLM-5.3-Flash launch post and the model overview, including the "Serving at Scale on Chinese AI Chips" section
• zai-org/GLM-5.3-Flash on Hugging Face (weights, MIT, config.json)
• OpenCode public model-usage dashboard (read 2026-08-27, 11:55 UTC)
• OpenRouter catalog entry and provider list
• South China Morning Post on the 62T preview total and the cluster-size figure