GLM-5.3-Flash
The mystery model topping the charts was Zhipu all along — and the reveal comes with an unmodified MIT file
What it is
For a week, an anonymous model called 'ox-alpha' quietly became the most popular model on OpenCode and OpenRouter — running, notably, on Chinese AI chips. On August 26 Z.AI claimed it: GLM-5.3-Flash, a 320B-total / 18B-active multimodal MoE mixing sparse and linear attention, taking text and images in, with day-one support across vLLM, SGLang, KTransformers, and Unsloth. Priced at roughly a tenth of GLM-5.2, positioned squarely at agentic coding.
Why it's interesting
The stealth-launch theater is fun, but the license is why it leads this week: the LICENSE file in the repo root is the standard, unmodified MIT text — 1.07 kB, 'Copyright (c) 2026 Z.AI Co., Ltd', no rider, no threshold, read directly on 2026-08-31. GLM-5.2 sits in our catalog already; 5.3-Flash beats it soundly on Z.AI's own tables (63.4 DeepSWE v1.1 vs 46.2, 84.3 Terminal-Bench 2.1) while activating fewer parameters. The uptake — ~379k downloads in week one — suggests the ox-alpha audience followed it home.
Use cases
- Self-hosted agentic coding at a fraction of flagship activation cost — 18B active per token
- Multimodal agent loops that read screenshots and UIs, MIT-licensed end to end
- Fine-tuning a current-generation hybrid-attention MoE without license negotiation
Who it's for
Agent builders, GLM-5.2 users due an upgrade, and anyone who was already running ox-alpha
Setup
Advanced. ~328GB of BF16 weights across the shard set; vLLM, SGLang, KTransformers, and Unsloth supported at launch, with community GGUF quants circulating
Limitations & cautions
Launch benchmarks are Z.AI's own and the ox-alpha popularity contest measured vibes, not rigor; the 1M-token context headline comes from coverage while the card's own evaluations run at 300K; and 320B total still wants serious hardware before quantization, whatever the activation count says.
Editorial takeaway
A frontier lab ran a blind taste test, won it, and then published the recipe under MIT. The stealth launch proved the model; the license file proves the policy.