GLM-5.2
A million tokens of context under plain MIT
What it is
Z.ai's 753B-total / ~40B-active MoE flagship with a 1M-token context window and IndexShare attention that cuts per-token FLOPs 2.9× at full length. Ships in BF16 and FP8 on Hugging Face with day-0 vLLM and SGLang support, plus llama.cpp, Ollama, and LM Studio quantizations.
Why it's interesting
The long-horizon workhorse of this summer's open-weights wave — MIT weights with no rider, no territory clause, no MAU cap. Already the engine behind community projects like Colibri, which streams its experts off disk to run it in ~25GB of RAM.
Use cases
- Whole-repository coding agents with 1M-token context
- Long-document analysis and sustained agentic sessions
- Self-hosted frontier-class assistant backends
Who it's for
Infra-rich teams; determined locals via community quantizations
Setup
Advanced. Multi-GPU node for BF16/FP8; community GGUF routes (see Colibri) trade latency for ~25GB-RAM consumer setups
Limitations & cautions
MoE inference at this scale is hardware-hungry, several hosts cap context well below 1M, and Z.ai's launch benchmark suite was thin — independent verification is still filling in.
Editorial takeaway
Colibri proved people will stream experts off a hard drive to run this model. A clean MIT license is why they bother.