On August 26, Zhipu pulled the mask off Ox Alpha, the model that had been running anonymously on OpenCode and OpenRouter for over a week. Its official name is GLM-5.3-Flash, and the weights went up on Hugging Face the same day under an MIT license.
It's the first natively multimodal member of the GLM-5 family. Total parameters stand at 320B, with only 18B active per token; the layer count drops from GLM-5.3's 92 to 45, and active parameters fall from the previous generation's 32B to 18B. The saved compute shows up directly in the price tag: the official list price is roughly one-tenth of GLM-5.3's, cut in half again during a limited-time promotion — working out to roughly one-fortieth of Claude Opus 4.8. Z.ai cites another figure: at the discounted rate, completing a task costs $0.045 on average.
Parameters cut in half, scores didn't follow
On Artificial Analysis's Intelligence Index v4.1.1, GLM-5.3-Flash scores 57, putting it in the same bracket as global frontier models — the same tier as Anthropic's currently most popular model, Opus 4.8. The spread across individual benchmarks is wider: DeepSWE v1.1 climbs from GLM-5.2's 46.2 to 63.4, Terminal-Bench 2.1 hits 84.3, Toolathlon Verified reaches 78.4, and AutomationBench v1.0.6 comes in at 48.8. On Zhipu's own Z.ai Code Bench, the hardest tier gives GLM-5.3-Flash a 29.0, against Opus's 29.5.
Multimodal capability is new this generation: OfficeQA Pro scores 62.4, CharXiv with tool-assisted reasoning hits 89.4, Chartography with tools reaches 78.0, and on the video side, MVBench scores 77.8 and MMVU 80.5. The combination points toward "an office agent that can read spreadsheets and diagrams," rather than a general image-understanding benchmark showcase.
Architecturally, it's the first open frontier model to mix sparse attention with linear attention, while also introducing IndexPool and mHC (manifold-constrained hyper-connections). The payoff shows up in two numbers: compared with GLM-5.3, attention compute drops to roughly a third (a 3.0x reduction) and the KV cache shrinks to roughly a quarter (4.4x). KV cache is the biggest cost driver in long-context workloads, and that multiplier determines whether a 1-million-token window is actually usable — not just a number on a spec sheet.
All the traffic runs on domestic chips
During the anonymous testing period, Ox Alpha was at one point the most-called model on OpenCode. Zhipu says all inference traffic during that stretch ran on domestic AI chips, with end-to-end serving performance up 3x compared with before, and costs now on par with mainstream Nvidia GPU setups.
That claim carries as much weight as the benchmark scores. The prevailing narrative for Chinese open models over the past year has been "scores are close, but deployment still leans on H-series cards." Once inference can run at full capacity on domestic silicon at matching unit cost, a model vendor's pricing is no longer hostage to GPU procurement prices. GLM-5.3-Flash's confidence in pricing at a tenth of the previous generation rests partly on this.
A rough calculation: at the limited-time rate of $0.075 per million input tokens and $0.25 per million output tokens, a mid-sized agent application averaging 50 million input tokens and 10 million output tokens a day would run up a model bill of about $6.25 a day (a rough figure, excluding cache discounts and retries). The same call volume on Opus 4.8 would differ by an order of magnitude. For teams building agent products in China, this pulls the question of "can we put this in production" back from budgeting into engineering.
On the deployment side, SGLang, vLLM, and KTransformers already support it, and the API is live simultaneously on docs.z.ai. The MIT license is an unusual move this time around — earlier GLM generations mostly shipped under Zhipu's own license, while MIT places almost no extra conditions on commercial redistribution, letting overseas teams modify it without another round of legal review.
A week earlier, Zhipu had already confirmed publicly that Ox Alpha was its own iteration, citing call volume more than double DeepSeek's at the time. Now the name, the weights, and the price are all on the table — what's left is how many teams are willing to move it from "cheap test model" to production workhorse.
Sources: Z.ai official blog, Hugging Face model page, CocoLoop, Artificial Analysis; parameter counts, benchmark scores, and API pricing verified against official release pages.