GLM-5.1

2026-04-07 · Publisher : Z.ai · Domain : Artificial Intelligence

Z.ai a ouvert GLM-5.1 sous licence MIT, modèle phare pour l'ingénierie agentique : 58,4 sur SWE-Bench Pro, et une capacité à rester efficace sur des centaines de tours et des milliers d'appels d'outils.

Z.ai has opened GLM-5.1 under the MIT license, a flagship model for agentic engineering: 58.4 on SWE-Bench Pro, and the ability to stay effective over hundreds of turns and thousands of tool calls. ### Context GLM-5 (February 2026) had set the 744B base. GLM-5.1 keeps it and mainly changes the post-training. According to Wikipedia, a subscriber-only version had been released in late March 2026, with the weights opened on April 7; the vendor then compares its model with Claude Opus 4.6, GPT-5.4 and Gemini 3.1 Pro. ### What the model brings - 58.4 on SWE-Bench Pro, the best score in that row of the published table (Opus 4.6: 57.3; GPT-5.4: 57.7; Gemini 3.1 Pro: 54.2). - NL2Repo (repository generation) 42.7 versus 35.9 for GLM-5; Terminal-Bench 2.0 (Terminus-2) 63.5 versus 56.2. - Better endurance over long sessions: previous models reach a plateau, while GLM-5.1 keeps improving, according to the vendor. - Example cited by the vendor and picked up by Computerworld: optimization of a vector database over more than 600 iterations and 6,000 tool calls, up to 21,500 queries per second. - The documentation release notes claim up to 8 hours of continuous execution and a capability alignment with Claude Opus 4.6. ### Architecture and specifications - Same architecture as GLM-5 (MoE, DSA): 744B announced for GLM-5; about 753.9 billion parameters by the HF count, identical to GLM-5. The number of active parameters is not restated in the 5.1 model card. - Context: 200K; maximum output: 128K; modality: text; knowledge cutoff: not disclosed. - License: MIT. Weights: zai-org/GLM-5.1 and GLM-5.1-FP8. - Deployment: SGLang, vLLM, xLLM, Transformers, KTransformers. ### Cost and access | Model (Z.ai API, current pricing) | Input /M | Cached input /M | Output /M | |---|---|---|---| | GLM-5.1 | 1.4 USD | 0.26 USD | 4.4 USD | ### Benchmarks | Benchmark | GLM-5.1 | GLM-5 | Claude Opus 4.6 | Gemini 3.1 Pro | GPT-5.4 | |---|---|---|---|---|---| | SWE-Bench Pro | 58.4 | 55.1 | 57.3 | 54.2 | 57.7 | | NL2Repo | 42.7 | 35.9 | 49.8 | 33.4 | 41.3 | | Terminal-Bench 2.0 (Terminus-2) | 63.5 | 56.2 | 65.4 | 68.5 | - | | CyberGym | 68.7 | 48.3 | 66.6 | 38.8 | 66.3 | | GPQA-Diamond | 86.2 | 86.0 | 91.3 | 94.3 | 92.0 | | HLE (text only) | 31.0 | 30.5 | 36.7 | 45.0 | 39.8 | | MCP-Atlas (public set) | 71.8 | 69.2 | 73.8 | 69.2 | 67.2 | Source: GLM-5.1 Hugging Face model card, scores published and measured by the vendor. ### Versus closed models On SWE-Bench Pro, GLM-5.1 scores 58.4, ahead of Claude Opus 4.6 (57.3), GPT-5.4 (57.7) and Gemini 3.1 Pro (54.2) in the vendor's table, with gaps of at most 1.1 points against the first two. It also leads on CyberGym (68.7 versus 66.6 for Opus 4.6). It remains behind on pure reasoning: 86.2 versus 94.3 for Gemini 3.1 Pro on GPQA-Diamond, and 31.0 versus 45.0 on HLE. On NL2Repo, Opus 4.6 is clearly ahead (49.8 versus 42.7). The API costs 1.4 USD for input and 4.4 USD for output per million tokens (current pricing); no contemporary closed-model pricing was verified to quantify the gap. The open choice is solid for long-horizon agentic code, less so for pure scientific reasoning tasks. ### Key takeaways GLM-5.1 shifts the emphasis from cold performance to endurance on long tasks, with an open model in the lead on SWE-Bench Pro according to the vendor. It keeps the MIT license and the architecture of GLM-5.

Tags : LLM, Open Source, Z.ai, Zhipu AI, GLM, MIT, MoE, Agents