Z.ai's GLM-5.3-Flash Is the First Natively Multimodal GLM — and the Weights Are MIT

For a week in late August, the most-used coding model on OpenRouter was a model nobody could name. It went by ox-alpha, it was free, it was fast, and it burned through more than 11 trillion tokens in its first three days on the platform — the biggest launch OpenRouter had ever recorded. On Wednesday, 26 August, Z.ai pulled off the mask. Ox Alpha was GLM-5.3-Flash, the first natively multimodal model in the GLM-5 family, and the company put the weights on Hugging Face under the MIT licence the same day.

The specifications check out against the primary sources. Z.ai's model card describes GLM-5.3-Flash as having "320B total parameters and just 18B active parameters" in a mixture-of-experts configuration — the Hugging Face safetensors index reports 321B, the usual rounding gap between marketing and tensor counts. The context window is 1 million tokens, with output capped at 128K on the hosted API. The licence field on the repository reads, simply, `mit`. All three headline claims in the original brief survive verification.

What is genuinely new is the architecture. GLM-5.3-Flash is not a fine-tune of GLM-5.3; it starts from a newly trained base model built on a hybrid of linear and sparse attention. "Linear attention captures local dependencies through state modeling, while sparse attention retrieves relevant global context through a lightweight indexer," the GLM team writes in its release notes. The company reports the design cuts attention compute by roughly 3x and shrinks KV cache size by 4.4x versus the full GLM-5.3 — a meaningful saving when you are quoting a million-token window. Z.ai also introduces Manifold-Constrained Hyper-Connections (mHC), and says the model was pretrained on a 30-trillion-token multimodal corpus. Vision is architectural rather than bolted on: the model was trained to inspect rendered interfaces, gameplay footage and 3D output, then critique and revise its own work from that visual feedback.

The benchmarks, with the usual caveats

On Artificial Analysis's Intelligence Index v4.1.1, GLM-5.3-Flash scores 57, which puts it in the same band as GPT-5.6 Terra, Gemini 3.7 Flash and Qwen 3.8 — at a measured cost of $0.045 per task. On agentic and coding evaluations the picture is stronger: 84.3 on Terminal-Bench 2.1, 63.4 on DeepSWE v1.1 (against GLM-5.2's 46.2), and 48.8 on AutomationBench versus 26.2 for the previous generation. Z.ai's own framing is that the model "approaches Claude Opus 4.8 on coding and agentic benchmarks" at one-tenth the price of GLM-5.2.

Two caveats travel with those numbers. First, the harnesses differ wildly — DeepSWE was run at 400K context with a six-hour timeout, Terminal-Bench inside Claude Code 2.1.207 — so cross-model comparisons are only as good as the setup. Second, the model is verbose. "One caveat: the model is very chatty and burns a lot of tokens," notes Frederic Lardinois, senior editor for AI at The New Stack, who tested it at launch. He adds the obvious rejoinder: "Because the inference costs are so low, though, that's not too much of an issue."

Those costs are the actual story. Z.ai's list price is $0.15 per million input tokens and $0.50 per million output, with cached input at $0.03 per million — and a 50% launch promotion running to 9 September 2026 takes that to $0.075/$0.25/$0.015. OpenRouter reflects the same discounted rates. The company also folded GLM-5.3-Flash into its Coding Plan with three times the usable quota of GLM-5.3.

Why MIT matters more than the parameter count

The open-weight field in mid-2026 is crowded, but it is not uniformly open. DeepSeek's V4 line and Z.ai's own GLM releases ship under MIT. Alibaba's Qwen 3.5 and 3.6 families use Apache-2.0. But Moonshot's Kimi K3 carries a custom licence with additional obligations for large companies and AI service providers, and MiniMax M3 is non-commercial by default. Those distinctions rarely make headlines and almost always decide procurement reviews. MIT imposes no field-of-use restriction, no attribution-in-product requirement, and no revenue threshold that flips a startup into a negotiation. For a company deciding what to embed in a shipped product, that is not a footnote — it is the whole decision.

The parameter economics reinforce it. Kimi K3's 2.8T weights run to roughly 1.6TB and demand multi-node deployment; GLM-5.3-Flash's 18B active parameters make it tractable on far less iron, with day-one support across SGLang, vLLM, TokenSpeed, KTransformers, Transformers and Unsloth, plus 79 community quantisations already on the Hub. The model had been downloaded more than 441,000 times within a week.

Then there is the hardware claim, which is the part Washington will read most closely. Z.ai says the entire ox-alpha preview — reportedly around 100,000 chips, serving up to 100 trillion tokens a day — ran on domestic Chinese accelerators. "Compared with our initial baseline on the same hardware, we achieved a 3x improvement in end-to-end serving performance, reaching hardware efficiency and per-token cost comparable to mainstream NVIDIA GPUs," the company writes, concluding that "Chinese chips can support frontier-model inference efficiently and economically at scale." Z.ai has not named the vendors, and independent verification of the chip claim is still outstanding — at least one outlet has flagged it as unconfirmed. Zhipu's Hong Kong-listed shares closed up more than 12% the day after the reveal.

What to watch

Three things. Whether third-party evaluators reproduce the agentic scores outside Z.ai's harnesses — the gap between 57 on a general index and near-Opus performance on coding is exactly where benchmark specificity hides. Whether the 50% promotion ends on 9 September or quietly becomes the new list price, which would tell you a great deal about the true serving margin. And whether anyone independently confirms the domestic-silicon claim. If that holds, the interesting number in this release was never 320 billion. It was 100,000.