The mystery model that spent a week charming developers under the anonymous name Ox Alpha now has an official identity. China's Z.AI has launched GLM-5.3-Flash, a natively multimodal, open-weight model released under the permissive MIT license — and the company says the entire stealth preview run was served on Chinese-made AI chips. The launch closes one of the most closely watched guessing games in the AI industry this year. For background on how the story unfolded, see our latest AI developments.
What Z.ai actually shipped
GLM-5.3-Flash is a mixture-of-experts model with 320 billion total parameters and 18 billion active parameters — a notable drop from the 32 billion active parameters of its GLM-4.5 predecessor. It is the first natively multimodal model in the GLM-5 family, accepting text and image input, and it supports a context window reaching one million tokens, according to Z.ai's announcement.
The model was trained on a 30-trillion-token multimodal corpus, and its architecture mixes linear attention for local dependencies with sparse attention for global context. A component Z.ai calls IndexPool compresses groups of indexer key vectors to limit latency and memory use at long context lengths. The company reports three times less attention compute and a 4.4-fold reduction in KV cache size compared with GLM-5.3.
Benchmark claims and pricing
On the Artificial Analysis Intelligence Index v4.1.1, GLM-5.3-Flash scored 57 — a figure that places it at the top of Artificial Analysis's current model rankings, well above the comparison median of 18. Z.ai says it outperforms GLM-5.2 across reported coding and agentic tests at one-tenth the price, and approaches Anthropic's Claude Opus 4.8 on its internal coding benchmark. On DeepSWE v1.1, the model scored 63.4 versus 46.2 for GLM-5.2; on AutomationBench it reached 48.8 against 26.2. As TestingCatalog noted, differing evaluation harnesses, context limits, and generation settings mean cross-test comparisons depend heavily on each setup.
Pricing positions the model as aggressively cheap: $0.15 per million input tokens and $0.50 per million output tokens, with an 83% prompt-cache discount, according to Artificial Analysis. Z.ai cites a discounted cost of $0.045 per Intelligence Index task.
The Chinese chip angle
The most strategically significant detail is infrastructure, not intelligence. Before launch, the model ran anonymously as `ox-alpha` on OpenRouter and OpenCode starting August 20, and Z.ai says it became the most popular model of the week on those services — with traffic served entirely on domestic Chinese accelerators. To support that, Z.ai built an SGLang-based serving stack that separates encoding, prefill, and decoding, reporting a threefold gain in end-to-end serving performance across tens of thousands of Chinese AI chips.
That matters beyond one vendor. It is a public demonstration that a frontier-adjacent Chinese model can be served at scale without Nvidia hardware, a data point Western policymakers and chipmakers will not ignore. The New Stack summarized the release as cheap, good, and served on Chinese chips — three qualities rarely claimed together until now.
Open weights, real deployment
The weights are available on Hugging Face under the MIT license, with local deployment supported through SGLang, vLLM, and TokenSpeed. GLM Coding Plan subscribers get GLM-5.3-Flash live immediately with three times the usable quota of GLM-5.3, and its multimodal capabilities are exposed in ZCode through Browser Use and Computer Use integrations.
Visual reasoning is central to the release: Z.ai trained the model to inspect rendered interfaces, gameplay, and 3D output, then assess and revise its own work from visual feedback. The same approach extends to documents, spreadsheets, presentations, and dashboards — the bread-and-butter artifacts of office work, which is exactly where Qwen's rival release the same day also claims its biggest gains.
The open-weights decision matters for the ecosystem beyond hobbyists. MIT licensing is the most permissive common license in AI: commercial use, modification, and redistribution carry no copyleft obligations. Independent hosts can serve GLM-5.3-Flash on their own hardware, fine-tune it for verticals from legal review to industrial automation, and embed it in products without negotiating terms — an option closed off by API-only frontier models.
Why the stealth launch worked
The anonymous rollout gave Z.ai something marketing budgets cannot buy: brand-agnostic validation. Developers adopted Ox Alpha on its merits, without the framing of a Chinese lab versus American incumbents. By the time Bloomberg confirmed Z.AI as the maker and Business Insider declared the "mystery solved," the model had already topped community leaderboards on neutral ground.
The competitive pressure is now explicit. A multimodal model with a one-million-token context, MIT-licensed weights, and sub-dollar output pricing forces every incumbent to justify its price list. The Pareto frontier for price-versus-performance — the trade-off developers actually feel — has been redrawn, as widely followed AI commentator Andrew Curran put it on X.
The open question is durability: whether benchmark leads hold up under adversarial real-world use, and how quickly Western labs respond on price. Either way, the week of anonymous speculation ended with a named model, downloadable weights, and a serving fleet that runs no American silicon.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
