zai-org/GLM-5

GLM‑5 Series – Large‑scale, long‑context LLMs from Z.ai

What it is – A family of open‑weight large language models (GLM‑5, GLM‑5.1, GLM‑5.2, GLM‑5.3 and the efficiency‑focused GLM‑5.3‑Flash) released by the Z.ai/GLM team. The models are built for coding, cyber‑security, and long‑horizon agentic tasks and support up to 1 million token contexts.

Why it matters – Compared with the previous GLM‑4.x line, GLM‑5 models double the parameter count (up to 744 B total, 40 B active) and increase pre‑training data to 28.5 T tokens. The team reports state‑of‑the‑art results on a range of public benchmarks (Terminal‑Bench, SWE‑Bench, CyberGym, Vending Bench, etc.) and claims the best open‑source performance on reasoning, coding and long‑term planning.

Key technical highlights

• Long‑context capability – GLM‑5.2 introduces a solid 1 M‑token context window, enabling sustained work on very long documents or multi‑step tool‑use.
• Hybrid sparse/linear attention – GLM‑5.3‑Flash adds a new architecture that mixes sparse and dense attention, cutting serving cost while keeping precise long‑range reasoning.
• IndexShare – A shared indexer across four sparse‑attention layers that reduces FLOPs by ~2.9× at 1 M context length.
• Manifold‑Constrained Hyper‑Connections (mHC) – Used in the Flash variant to improve scaling efficiency.
• Reasoning‑effort control – The newer models expose a reasoning_effort parameter (low, high, max) and a clear_thinking flag to trade latency for performance.
• Reinforcement‑learning infrastructure – The slime library (asynchronous RL) is provided for fine‑tuning and post‑training.

Model sizes & formats

Model	Parameters (total / active)	Precision options	Where to download
GLM‑5.3	744 B / 40 B	FP8, BF16	Hugging Face, ModelScope
GLM‑5.3‑Flash	320 B / 18 B	FP8, BF16	Hugging Face, ModelScope
GLM‑5.2	744 B / 40 B	BF16, FP8	Hugging Face, ModelScope
GLM‑5.1	744 B / 40 B	BF16, FP8	Hugging Face, ModelScope
GLM‑5	744 B / 40 B	BF16, FP8	Hugging Face, ModelScope

Running the models locally – The repository lists several inference back‑ends that already have recipes for the GLM‑5 series:

• SGLang – fast serving with a dedicated cookbook.
• vLLM – high‑throughput inference; recipes are provided for each model.
• Transformers (🤗) – official model cards and docs.
• KTransformers – kernel‑level acceleration tutorials.
• Unsloth – lightweight deployment guide.
• Ascend NPU – support via vLLM‑Ascend, xLLM, etc.

Fine‑tuning – The models can be further adapted with:

• slime (the team’s async RL framework) for reinforcement‑learning‑based finetuning.
• ms‑swift for supervised fine‑tuning (SFT), PPO, and GRPO.

Community & resources

• Official Discord and WeChat groups for user support.
• Blog posts describing GLM‑5.3, GLM‑5.3‑Flash, and GLM‑5.2.
• Technical report on arXiv (arXiv:2602.15763) that details the architecture and training data.

Citation – If you use any GLM‑5 model in research, cite the provided arXiv technical report (full BibTeX included in the README).

---

All information above is taken directly from the repository’s README; no additional claims have been added.