GLM-5.3 Goes Open-Weight

Zhipu AI quietly dropped GLM-5.3 in open-weight form, and it's getting serious traction — 50k+ HuggingFace downloads, multiple stories crossing 1,000 points on HN today. The full-weight variant (GLM-5.3) uses a restrictive custom license, but the GLM-5.3-Flash variant is MIT-licensed and multimodal, already at 346k downloads.

The headline: every gain over GLM-5.2 comes from post-training alone. Same base architecture. Same MoE-DSA backbone. They scaled post-training and emergent capabilities appeared — especially in coding and cybersecurity.

On Terminal Bench 3.0, GLM-5.3 scores 28.3 vs. 4.6 for GLM-5.2 — a 6x improvement. On DeepSWE, it hits 66.9 (up from 46.2). On ExploitBench, it more than doubles from 24.4 to 54.4. The model is competitive with Kimi K3, DeepSeek-V4 Pro, and Opus 4.8 on most coding benchmarks, and beats them outright on CyberGym and AutomationBench.

What makes this interesting isn't just the benchmark table — it's what the numbers don't say. GLM-5.3 is an open-weight model trained in China, deployable on Ascend NPUs, that competes with closed-source frontier models. The cyber capability isn't a footnote — their own writeup calls it "emergent," saying it "developed faster than we expected" as they scaled post-training. Exploitation benchmarks more than doubled from 5.2 to 5.3.

Deployment-wise, it works with SGLang, vLLM, Transformers, KTransformers, and Unsloth out of the box. Supports a reasoning_effort parameter (low/high/max) and ships pre-quantized to FP8. The Flash variant is already the breakout — MIT license, multimodal, more downloads in days than most models see in months.

The open-weight frontier just got a lot more interesting. Keep an eye on the Flash variant — that's the one that actually ships permissively.