GLM-5.3-Flash is an open-weight reasoning model released by Z.ai on August 26, 2026. Z.ai describes it as "a reasoning model designed for coding, sustained agentic work, and production workloads" and as the newest iteration of its GLM series. It had circulated anonymously on OpenRouter as Ox Alpha from around August 20, 2026 before the company claimed it.
| Field | Value |
|---|---|
| Developer | Z.ai (Zhipu AI) |
| Released | August 26, 2026 |
| Preview name | Ox Alpha (anonymous on OpenRouter, from ~August 20, 2026) |
| Total parameters | 320 billion |
| Active parameters | 18 billion |
| Layers | 45 |
| Open weights | Yes, published on Hugging Face |
| Serving hardware | Chinese AI chips, dedicated SGLang inference engine |
Architecture
Z.ai's technical post gives 320 billion total parameters with 18 billion active, against 355 billion total and 32 billion active for GLM-4.5, and 45 layers against 92. The model is therefore smaller in every published dimension than the predecessor it is compared to — roughly half the active parameters and half the depth — while the company reports higher scores on its agentic and coding benchmarks. The "Flash" designation and the reported cost figures place it as a cost-efficiency release rather than a capability-ceiling one.
Reported benchmarks
All figures are Z.ai's own; no independent evaluation has been published.
| Benchmark | GLM-5.3-Flash | Comparison |
|---|---|---|
| DeepSWE v1.1 | 63.4 | 46.2 (GLM-5.2) |
| AutomationBench v1.0.6 | 48.8 | 26.2 (GLM-5.2) |
| Terminal Bench 2.1 | 84.3 | 85.0 (Claude Opus 4.8) |
| Artificial Analysis Intelligence Index v4.1.1 | 57 | at $0.045 per task, discounted |
The Terminal Bench result is the notable comparison: Z.ai reports the model within a point of Claude Opus 4.8 on that benchmark while claiming a per-task cost of $0.045 discounted.
Serving on domestic hardware
Z.ai states that all Ox Alpha traffic during the anonymous preview was served on a large-scale cluster of Chinese AI chips; that it built a dedicated inference engine on SGLang for that hardware; and that it reached a threefold end-to-end serving improvement over its initial baseline, at per-token cost comparable to mainstream Nvidia GPUs.
That claim is the release's most consequential, and its evidentiary status differs from the benchmark figures. The anonymous preview means the traffic was served to third-party users on OpenRouter for roughly a week before the hardware claim was made, so the latency and throughput characteristics were observable in public even though the underlying hardware was not. Z.ai does not name the chips. If the cost-parity claim holds, it bears on the premise of export controls — that restricting access to leading accelerators constrains Chinese frontier deployment — though a single vendor claim about one model's serving economics is not sufficient to establish it.
Anonymous preview
The model was previewed on OpenRouter from around August 20, 2026 without attribution and was unclaimed as of August 23. A fingerprinting tool matched its infrastructure signature to Z.ai's, which was read at the time as circumstantial evidence without establishing authorship; the company's August 26 confirmation resolved it. Early capability claims moved during the preview: developer Ben Davis first scored the model above 80% on ten DeepSWE tasks, against 65% for Claude Fable 5 and 52% for GPT-5.6 Sol, then ran the full DeepSWE set and found it roughly level with GPT-5.6 Sol mid rather than well ahead (Source: techcrunch.com; z.ai).
Weights are published on Hugging Face at zai-org/GLM-5.3-flash. Z.ai published no safety case or system card with the release, and the post does not address cyber capability — the ground on which the company had delayed the open-weight release of GLM-5.3 earlier the same month.
Open questions
- Whether independent evaluation reproduces the DeepSWE, AutomationBench, and Terminal Bench figures now that weights are available.
- Which Chinese AI chips served the preview, and whether the stated per-token cost parity with Nvidia hardware holds under third-party measurement.
- Why this release carries no cyber-capability assessment when GLM-5.3's open-weight release was delayed on that ground three weeks earlier.
- Whether releasing under an anonymous preview name becomes a repeated pattern for Chinese developers, and what it is intended to test.
Relationships
- related: GLM-5.3 — the same-month sibling release, whose open-weight publication was delayed on cyber-capability grounds.
- related: GLM-5.2, GLM-4 generation (GLM-4 through GLM-4.7) — the comparison points in Z.ai's own benchmark table.
- related: Zhipu AI — the developer.
- related: OpenRouter — the platform hosting the anonymous preview.
- related: Export Controls (AI) — the domestic-hardware serving claim bears on the controls' premise.
- related: Open-Weight Frontier Models.