What it does
GLM-5.1 is described by Z.ai as its next-generation flagship model for agentic engineering, with much stronger coding than its predecessor. The repository reports state-of-the-art results on SWE-Bench Pro and wide leads over GLM-5 on NL2Repo (repository generation) and Terminal-Bench 2.0 (real-world terminal tasks). The stated aim is to stay effective over long agentic sessions. Where earlier models plateau after quick early gains, the page says GLM-5.1 breaks complex problems down, runs experiments, reads results and revises its strategy, sustaining optimisation over hundreds of rounds and thousands of tool calls. The page is the README for the whole GLM-5 series. It lists GLM-5.1 weights at 744B-A40B in BF16 and FP8 on Hugging Face and ModelScope, serving guides for SGLang, vLLM, Transformers, KTransformers and Unsloth, Ascend NPU support through vLLM-Ascend, xLLM and SGLang, and fine-tuning with Slime and ms-swift. The same README also covers GLM-5.2, GLM-5.3 and GLM-5.3-Flash, and the repository carries an Apache-2.0 licence.
Capabilities
- Aimed at agentic engineering and coding workflows
- Reported state-of-the-art result on SWE-Bench Pro
- Built to keep working over hundreds of rounds and thousands of tool calls
- 744B-A40B weights in BF16 and FP8 on Hugging Face and ModelScope
- Serving guides for SGLang, vLLM, Transformers, KTransformers and Unsloth
- Fine-tuning with Slime and ms-swift
When to use it
- You want downloadable weights for a coding and agent model that you host yourself
- You are evaluating models for long-running agent tasks
- You plan to fine-tune a GLM-5 series model with reinforcement learning or SFT frameworks
