Z.ai opens GLM-5.3's full 753B weights — not MIT this time
Frontier-class MoE, ~40B active, downloadable now for vLLM/SGLang; the gains came from post-training, not a new base. Plus: a court voids the Anthropic ban.
Copy markdown
The full model, not the Flash
Z.ai released open weights for its flagship GLM-5.3 on Hugging Face — ~753B total, ~40B active per token, MoE, ~200K context. Safetensors are live now in BF16 and FP8, distinct from last week's smaller 320B GLM-5.3-Flash.
Post-training did the work
Z.ai says the base is unchanged from GLM-5.2; the jump came from roughly a month of extra RL on long-horizon tasks. Terminal-Bench 3.0 went 4.6 to 28.3 and DeepSWE 46.2 to 66.9, with vendor-claimed open-source SOTA on several agentic benchmarks (~60 on the AA Intelligence Index).
The license is the catch
Unlike GLM-5.2 and the Flash variant — both MIT — the flagship ships under a custom "glm-5.3" license, not MIT. Read the terms before you redistribute it or bake it into a product; don't assume GLM-5.2's permissions carry over.
How to run it this weekend
vLLM and SGLang support it day-one with OpenAI-compatible endpoints, so pointing your client at it is a base-URL change. Budget ~745GB VRAM at FP8 or ~370-400GB for INT4 community quants — multi-GPU only. No hardware? It's already served on OpenRouter.
Elsewhere: a court voids the Pentagon's Anthropic ban
A federal judge ruled the Pentagon's designation of Anthropic as a supply-chain risk unlawful, striking down the blacklist. If you ship Claude-based tools into government or defense workflows, a procurement door that was closing just reopened.