Z.ai · 2026-04-07 · major
GLM-5.1
Open-weight 754B MoE model from Z.ai focused on long-horizon agentic coding. Scores 58.4 on SWE-Bench Pro — ahead of GPT-5.4 and Opus 4.6, under an MIT license.
A 754B open-weight Mixture-of-Experts model from China that beats GPT-5.4 and Claude Opus 4.6 on SWE-Bench Pro — under an MIT license.
Key specs
| License | MIT |
|---|---|
| Parameters | 754B MoE |
| Swe bench pro | 58.4% |
What is it?
GLM-5.1 is the next-generation flagship model from Z.ai (formerly Zhipu), aimed squarely at agentic engineering — the workflow where a model reads an issue, plans fixes, writes code, runs tests, and iterates. It is the first open-weight model to top SWE-Bench Pro, scoring 58.4% versus 57.7% for GPT-5.4 and 57.3% for Claude Opus 4.6.
How does it work?
GLM-5.1 is a 754B-parameter Mixture-of-Experts model built on the GLM-5 base architecture with DeepSeek Sparse Attention. Z.ai specifically trained it for sustained work: the docs claim it can stay productive across hundreds of turns and thousands of tool calls, breaking a task down, running experiments, reading results and revising strategy. Weights ship on Hugging Face (BF16 and FP8) under an MIT license.
Why does it matter?
A frontier-grade open-weight coder under MIT is the cleanest challenge yet to closed-model pricing for agentic coding. Teams running long-horizon agents on Claude Code or OpenAI Codex can now self-host a model that is at least competitive on SWE-Bench Pro without the per-token bill.
Who is it for?
Self-hosters, agent framework authors, anyone building around Claude Code or Codex who wants an open fallback.
Try it
huggingface.co/zai-org/GLM-5.1