<p>ZhipuAI (<a href="http://Z.ai" target="_blank" rel="noopener nofollow">Z.ai</a>) has released GLM-5.1, an open-source model built specifically for long-running agentic software engineering. It’s available on Hugging Face and ModelScope under the MIT license, meaning you can download, fine-tune, and run it locally.</p>
<h2>Benchmark results</h2>
<p>GLM-5.1 scores 58.4 on SWE-Bench Pro, which puts it first among open-source models and third globally — ahead of GPT-4 (57.7) and Claude Opus 4.6 (57.3) on that benchmark. It also leads on NL2Repo and Terminal-Bench 2.0. On Code Arena it ranks third, behind Opus 4.6.</p>
<p>The usual caveats apply: public benchmarks don’t always reflect messy real-world codebases, so treat these numbers as directional rather than definitive.</p>
<h2>What makes it different</h2>
<p>Most coding models degrade after a few dozen turns. GLM-5.1 is designed to keep improving over hundreds of iterations and thousands of tool calls — the kind of sustained autonomous work that multi-hour agentic tasks require.</p>
<p>A few concrete examples from ZhipuAI’s release:</p>
<ul>
<li><strong>Vector search optimization (VectorDBBench):</strong> Over 600+ iterations and roughly 6,000 tool calls, the model reached 21,500 QPS — about 6x better than what a single 50-turn session produced (3,547 QPS).</li>
<li><strong>GPU kernel optimization (KernelBench Level 3):</strong> Achieved a 3.6x speedup.</li>
<li><strong>Open-ended web app build:</strong> After 8 hours of self-directed iteration, it produced a full Linux-style desktop environment running in the browser.</li>
</ul>
<h2>Availability and compatibility</h2>
<p>Weights are on Hugging Face and ModelScope. The model works with Claude Code, Cline, and other standard coding agent frameworks. It’s also available in HuggingChat and LlamaCoder for quick testing without local setup.</p>
<p>For enterprises in regulated industries where running models on your own infrastructure matters, the MIT license and local deployment option are the practical draw here — not just the benchmark numbers.</p>


Collections

ZhipuAI has released GLM-5.1, an open-source coding model under the MIT license available on Hugging Face and ModelScope. It scores 58.4 on SWE-Bench Pro, ranking first among open-source models and third globally, ahead of GPT-4 and Claude Opus 4.6. Unlike typical coding models that degrade after dozens of turns, GLM-5.1 is designed for sustained agentic work over hundreds of iterations and thousands of tool calls. Demonstrated results include a 6x QPS improvement in vector search optimization over 600+ iterations, a 3.6x GPU kernel speedup, and an 8-hour autonomous build of a browser-based Linux desktop environment. It integrates with Claude Code, Cline, and other agent frameworks, and supports local deployment — making it particularly relevant for enterprises in regulated industries.

GLM-5.1 is now on Hugging Face, ranking #1 in open source on SWE-Bench Pro