For six days in late August, the most talked-about model on OpenRouter had no name and no owner. It appeared as stealth/ox-alpha, offered a one-million-token context window, cost nothing to use, and climbed straight to the top of the leaderboard. Developers ran forensics on it. Stack traces, an obscure error code, and tokenizer comparisons all pointed one direction.
They were right. Ox Alpha belonged to Z.ai, the Chinese lab behind the GLM series. Bloomberg reported the connection on the morning of August 26. That evening, Z.ai confirmed it and gave the model its production name: GLM-5.3-Flash.
Weights went up on Hugging Face the same night under an MIT license. API documentation went live alongside them. The stealth listing had been a preview skin for a product that was already finished.
The specifications behind the leaderboard run
GLM-5.3-Flash is a sparse mixture-of-experts design. It carries 320 billion total parameters but activates roughly 18 billion per token. That ratio is the entire economic argument for the model.
It is natively multimodal across text, image, and video rather than a text model with vision attached afterward. Z.ai frames vision as part of the coding loop, where a model renders output, checks it, and refines. The context target is one million tokens, supported by a hybrid linear and sparse attention scheme.
The company claims that architecture cuts serving cost substantially against its own larger GLM-5.3, citing roughly three times less attention compute and a KV cache about 4.4 times smaller on per-token metrics. Pre-training ran on a multimodal corpus the lab describes as 30 trillion tokens.
Pricing is where the pressure lands. List rates are 15 cents per million input tokens and 50 cents per million output tokens, with cached input at 3 cents. Those are Flash-tier numbers for a model being positioned against frontier systems.
Reading the benchmark table carefully
Z.ai published an evaluation table at launch, and it deserves a skeptical eye because the vendor produced it. The results are also not a clean sweep.
The clearest gains are against the lab's own previous generation. On DeepSWE the model reports 63.4 against 46.2 for GLM-5.2. On AutomationBench it reports 48.8 against 26.2. Both are agentic workloads, which matches the traffic Ox Alpha was absorbing on OpenRouter.
Against outside competition the picture is mixed. Z.ai reports leading GDPVal-AA v2, placing it ahead of Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash on that measure. But GPT-5.6 Terra still leads on DeepSWE and Terminal Bench in the same table, and Gemini 3.7 Flash leads AutomationBench.
One widely shared figure from the stealth period, an 80 percent DeepSWE score, sits far above the official 63.4. The gap most likely reflects a different evaluation harness, which is a useful reminder that agent benchmarks are not portable between setups.
The line that will get quoted
Z.ai added one detail to its announcement that carries more weight than any benchmark row. The entire stealth preview, it said, was served on Chinese AI chips.
That claim reframes the release. It is no longer only a story about a cheap capable model. It is a claim about domestic inference capacity at leaderboard-topping scale, made at a moment when export controls are meant to constrain exactly that.
The commercial threat to expensive frontier vendors is straightforward. Open weights under MIT terms let teams self-host and fine-tune without a licensing conversation. Launch-day support for SGLang, vLLM, and TokenSpeed removes most of the integration friction. A sub-dollar output price removes the rest.
The playbook itself is worth noting, because Z.ai has now demonstrated it works. Ship anonymously, let a developer community validate the model without brand bias, become the most used model of the week, then attach the name and release the weights. Earned credibility arrives before the marketing does.
Anyone building on the preview should migrate off the stealth identifier and route to the named model or a self-hosted copy, since preview endpoints are not guaranteed to persist.
