Z.ai (Zhipu) has officially launched GLM-5.3-Flash, and it is the model coders spent the past week chasing under a codename. The stealth "Ox Alpha" that dominated OpenRouter's free-model usage charts and ended DeepSeek's 56-day run atop the OpenCode leaderboard turns out to be this release. GLM-5.3-Flash ships under an MIT license with open weights, a 320-billion-parameter mixture-of-experts design, and a 1-million-token context window.

Try It: Route a Coding Agent to GLM-5.3-Flash Today

Point any OpenRouter-compatible tool at GLM-5.3-Flash and run a real task. In an agent like OpenCode, Cursor, or a custom script, set the provider to OpenRouter or the Z.ai API and select GLM-5.3-Flash, then fire a multi-file refactor, a repo-wide bug hunt, or a long-context review that loads your entire codebase into the 1M window. At $0.15 per million input tokens and $0.50 per million output tokens, it is cheap enough to benchmark head-to-head against your current coding model on your own work before you commit.

Why It Matters

A top-of-leaderboard coding model arriving as MIT-licensed open weights resets what self-hosted builders can run. Z.ai already shipped GLM-5.3 as an open-weight coding model, and the Flash variant pushes that further: comparable coding and agentic scores at roughly one-tenth the cost of GLM-5.2, with weights on Hugging Face for local deployment via SGLang, vLLM, or TokenSpeed.

Key Details

Architecture: 320B total parameters, 18B active (mixture-of-experts), down from 32B active in GLM-4.5.

Context: 1M tokens. License: MIT. Pricing: $0.15 input / $0.50 output per million tokens.

Benchmarks: Artificial Analysis scores it at an Intelligence Index of 57, with DeepSWE v1.1 at 63.4 and AutomationBench at 48.8. Treat the viral "beats GPT-5.6 Sol and Claude Fable 5" claim with caution: it came from an informal 10-task run, and other tests such as LiveBench rank it lower. Z.ai confirmed the MIT release and the Ox Alpha connection at launch.

What to Do Next

Add GLM-5.3-Flash to your provider list and run it against the same tasks you give your paid coding model, watching how it handles your largest-context jobs. If it holds up, self-hosting the weights removes the per-token cost entirely. For context on how it stacks against the other open flagship that held the top spot before it, see our coverage of DeepSeek V4-Pro.