GLM-5.3-Flash Is Ox Alpha: 320B, Multimodal, MIT

Z.ai released GLM-5.3-Flash on August 26, and in doing so ended a week of guesswork: the model is Ox Alpha, the anonymous system that had been quietly serving free requests on OpenRouter and in the OpenCode client since August 20. OpenRouter now says so directly on its model page, where the listing notes that the stealth model was revealed to be Z.ai’s GLM-5.3-Flash.

When we wrote about Ox Alpha two days ago, the leading theory already pointed at Z.ai — tokenizer probing and behavioural fingerprinting kept landing on the GLM family. The objection was that Ox Alpha accepted images and video while GLM-5.3 is text-only. GLM-5.3-Flash resolves that cleanly: it is not GLM-5.3 with a vision bolt-on but a separate, natively multimodal model that takes text, images and video and returns text.

Architecturally it is a mixture-of-experts model with 320 billion total parameters and 18 billion active per token, which is what makes the “Flash” positioning credible. Z.ai attributes the efficiency to three changes: a hybrid attention stack that interleaves KDA linear-attention layers with NoPE sparse MLA layers, which the company says cuts attention compute roughly threefold and shrinks the KV cache by 4.4× against GLM-5.3; an IndexPool step that compresses indexer key vectors by weighted pooling to hold latency down at million-token scale; and manifold-constrained hyper-connections (mHC), which roughly halve both activated parameters and layer count relative to GLM-4.5. Context is listed at 1,048,576 tokens with up to 131,072 tokens per completion, though OpenRouter’s own page quotes a slightly higher 1,310,720-token ceiling.

The benchmark numbers Z.ai published are self-reported, and they are aimed squarely at the top of the coding leaderboards. GLM-5.3-Flash scores 84.3 on Terminal-Bench 2.1 against 85.0 for Claude Opus 4.8, and 29.0 on Z.ai’s own code benchmark against Opus 4.8’s 29.5 — within half a point in both cases. Against its own predecessor the jumps are larger: 63.4 versus 46.2 on DeepSWE v1.1, and 48.8 versus 26.2 on AutomationBench. Independently, Artificial Analysis puts it at 57 on its Intelligence Index, well clear of the 27 median for open-weight models in the same size class.

Pricing is the other half of the pitch. Z.ai lists $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens, with a launch promotion halving those rates through September 9. The weights are MIT-licensed and published on Hugging Face, at roughly 306 GiB in FP8 — which means self-hosting needs Hopper-generation NVIDIA silicon or newer. Subscribers to the GLM Coding Plan tiers get about three times the usable quota they had on GLM-5.3.

The detail most likely to travel furthest has nothing to do with benchmarks. MarkTechPost’s writeup of the release reports that the Ox Alpha preview was served entirely on domestically produced Chinese AI accelerators. If that holds up under scrutiny, the anonymous free-tier rollout reads less like a marketing stunt and more like a load test — a way to put a frontier-adjacent coding model in front of tens of thousands of developers, on non-NVIDIA hardware, before anyone knew whose reputation was on the line.