GLM-5.3 open weights drop: 744B MoE, permissive custom license
The flagship weights land on Hugging Face ~2.5 weeks post-launch: GGUF quants run on a 256GB box, and commercial use is free unless you're a $10B cloud.

Copy markdown
The weights are actually live now
Z.ai's own org shipped zai-org/GLM-5.3 to Hugging Face — roughly 744B total parameters with ~40B active (MoE), in BF16 and FP8 safetensors. It lands about 2.5 weeks after the Aug 14 coding-service launch, once the promised security review cleared. Anything not on the zai-org org isn't the real model.
The license is new — and fine for almost everyone
This is the first GLM to ship under a custom “GLM-5.3 License” rather than the MIT terms that covered 5.1 and 5.2. It still grants rights to use, modify, distribute and sell; the one catch is that a Model-as-a-Service business clearing $10B in revenue over any 12 months must pass Z.ai's security review first. For solo devs and startups, that's effectively unrestricted commercial use.
You can run it on a workstation, not a rack
Unsloth's dynamic GGUFs are already published: the 2-bit UD-IQ2_M is ~239GB and runs on a 256GB-RAM machine, and a 1-bit build drops to ~223GB. Full BF16 still wants ~810GB, but a maxed-out Mac Studio or big-RAM box now handles the flagship locally.
Strong at coding and cyber-defense, not a frontier-killer
Z.ai claims ~50% over GLM-5.2 on its own Code Bench and 84.5% on CyberGym defense, edging Fable 5. But it trails the West on general coding — 66.9% on DeepSWE vs GPT-5.6 Sol's 72.7% — so treat it as a top-tier open coding and agent model, not a GPT-5.6 replacement.
Prefer an API? It's OpenAI-compatible
If running 744B locally isn't your weekend, the hosted model is served on an OpenAI-compatible endpoint plus the GLM Coding Plan (Lite $12.6/mo, Pro $56, Max $117.6), with a reasoning_effort switch (low/high/max) to trade latency and cost for depth.
Elsewhere: Claude Code 2.1.252 fixes Mac Bash failures
A same-day point release quietly fixes Bash commands dying with “task output swap refused” on some Macs, “always allow” not saving in projects without a local settings file, and Remote Control sessions stalling on flaky connections.