Z.ai released GLM-5.3-Flash on August 26 and confirmed it is the model that had been running anonymously as "ox-alpha" on OpenRouter and OpenCode, where it became the most popular model of the week. The weights are public under the MIT licence: 320 billion total parameters with 18 billion active, natively multimodal, with a one-million-token context window. The detail with the longest reach is not the benchmark -- it is that z.ai says it served all of that anonymous traffic on Chinese AI accelerators.
Key facts
- 320B total parameters, 18B active, 45 layers, 288 routed experts plus one shared, 1,048,576-token context.
- Released August 26, 2026, under the MIT licence, by Z.ai (formerly Zhipu AI).
- The download is 328 GB of already fp8-quantized weights across 72 files.
- Primary source: Z.ai, "GLM-5.3-Flash: Frontier Intelligence, Flash Cost"; weights at
zai-org/GLM-5.3-Flash.
The hook. For a week, the best-value model on OpenRouter had no name and no parent. Practitioners called it Ox Alpha and argued about who made it. The answer arrived with the licence file attached.
Background. Z.ai is the Beijing lab behind the GLM series; we covered GLM-5.3 shipping with a ledger of 2,436 security findings earlier this month. Testing a model anonymously on a public router before release is now a common tactic: it gets you real usage data without brand halo or brand suspicion contaminating the feedback.
What they did. GLM-5.3-Flash is the first natively multimodal model in the GLM-5 line, trained on a 30-trillion-token multimodal corpus. The architecture is the interesting part, and the configuration file on Hugging Face confirms it directly: of 45 layers, 34 use linear attention and 11 use DeepSeek-style sparse attention. Compared with the GLM-4.5 generation it holds roughly the same total parameter count (320B against 355B) while nearly halving both the active parameters (18B against 32B) and the layer count (45 against 92). Against GLM-5.3 it cuts attention compute by a factor of 3.0 and KV cache size by 4.4.
How it works. Attention is the mechanism that lets a model relate every word to every other word, and its cost grows brutally as the context gets longer. Linear attention trades that for a running summary -- cheap, good at local detail, weaker at reaching far back. Sparse attention keeps the full-strength version but only for a small selected subset of the context, chosen by a lightweight "indexer." Alternating them is like a reader who skims most pages fast and stops to read closely on the few that matter. Z.ai adds a compression trick it calls IndexPool, which squeezes four indexer key vectors into one by weighted pooling, specifically to keep the indexer affordable at a million tokens. The model is also a mixture of experts: 288 specialists exist, eight run per token.
Why it matters. Z.ai reports pushing the frontier of the Artificial Analysis Intelligence Index at roughly one-tenth the cost of models at comparable capability -- and the company is explicit about how it got there. It built a dedicated inference engine on top of SGLang for domestic hardware, using W8A8 quantization, hybrid cache quantization, layer split, and a production Encode-Prefill-Decode architecture that separates multimodal encoding, prompt prefill and token-by-token decoding into independently scaled worker pools "across tens of thousands of domestically developed accelerators." The result, in z.ai's words: "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. This demonstrates that Chinese chips can support frontier-model inference efficiently and economically at scale."
There is a recursive detail buried in that section that is easy to miss. Z.ai says the serving stack was built with help from "our GLM-5.3-powered infrastructure agent, which assisted engineers in developing and optimizing kernels, diagnosing performance bottlenecks, and improving the serving stack -- creating a feedback loop in which the model helped optimize the system serving the model itself."
The honest caveat. Three, actually. Every benchmark figure in the release is z.ai's own, including an in-house coding evaluation, so the comparisons against Claude Opus 4.8 are vendor numbers until someone retests them. The cost-parity-with-NVIDIA claim is measured against z.ai's own earlier baseline on the same domestic hardware, not against an NVIDIA cluster running the same model. And 328 GB of weights is not something you run at home -- open weights increasingly means "auditable and portable," not "runnable on your desk." The free Ox Alpha window is also over; the model is now a paid product.
Still, MIT is MIT. A 320-billion-parameter multimodal model with a million-token context, released with no field-of-use restriction at all, is the most permissive frontier-adjacent release of the month -- and it lands the same week Qwen shipped its own open-weight flagship with a considerably more restrictive contract.
Originally published on Ground Truth, where every claim is checked against the primary source.
