GLM-5.1 Goes Open-Source and Hits #1 on SWE-Bench Pro — Beating Every Closed AI Model
Z.ai's GLM-5.1 is a 754B open-weight MoE model under the MIT license — and it just took #1 on SWE-Bench Pro, outscoring every major closed model.
The First Open-Source Model to Beat All Closed Frontier Models on Real-World Coding
April 7, 2026 produced one of the more striking benchmark results in recent open-source AI history: GLM-5.1, released by Z.ai (formerly Zhipu AI) under the MIT license, posted a 58.4 score on SWE-Bench Pro — placing it at the top of the global leaderboard for software engineering AI, ahead of every major closed-source frontier model currently on the market.
The numbers are straightforward: GPT-5.4 scored 57.7, Claude Opus 4.6 scored 57.3, Gemini 3.1 Pro scored 54.2. GLM-5.1 leads all of them.
This is the first time an open-weight model has taken the top position on a real-world software engineering benchmark that the industry broadly treats as a credible signal of practical coding capability. That's a significant milestone for the open-source AI ecosystem.
Architecture: Scale Without the Inference Cost
GLM-5.1 uses a Mixture-of-Experts architecture designed to be practical to run, not just impressive on paper. Total parameter count is 754 billion, but only approximately 40 billion parameters activate per inference — keeping computational costs manageable while accessing the full model's representational capacity for any given task.
The context window extends to 200K tokens, which is directly relevant to the model's design target: long-horizon agentic coding tasks that require keeping a full codebase, test suite, and problem definition in active context simultaneously.
Built for Agentic, Long-Running Software Engineering
GLM-5.1 is specifically engineered for the kind of extended autonomous software engineering workflows that current AI coding tools begin to struggle with at scale. Z.ai published benchmarks showing the model maintaining coherent coding strategy across 600+ consecutive iterations — a capability that directly addresses one of the core failure modes of AI software agents on complex real-world projects.
The distinction matters: SWE-Bench Pro measures performance on real GitHub issues requiring understanding of existing codebases, identifying the right files, writing code changes, and passing test suites. Reaching #1 on that benchmark with an open-weight model, available under the most permissive commercial license available, is a genuinely useful development for the developer ecosystem.
MIT License: The Permissive Commercial Choice
The MIT license that Z.ai chose allows commercial use, closed-source derivative products, and secondary fine-tuning with no restrictions beyond retaining the copyright notice. Developers can integrate GLM-5.1 directly into commercial SaaS products, private enterprise deployments, and fine-tuning pipelines without licensing fees or open-source obligations.
The model weights are available now on Hugging Face at zai-org/GLM-5. Developers can run it today.
What This Means for AI Software Development
GLM-5.1 arriving at #1 on SWE-Bench Pro reinforces a pattern that's been building in open-source AI: the gap between open-weight and closed frontier models continues to compress. For teams building AI-assisted software engineering workflows, having a top-performing agentic coding model available to run on-premise — under a commercial-friendly license — is a meaningful expansion of what's practical without API dependency.
Sources: TechBriefly (April 8, 2026), The Decoder (April 2026), Z.ai GLM-5.1 Hugging Face release (April 7, 2026), WhatLLM (April 2026)