Krux
April 9, 2026
Z.ai's GLM-5.1 Can Code Autonomously for Hours
Published: April 9, 2026 at 12:35 AM
Updated: April 9, 2026 at 12:35 AM
100-word summary
Z.ai released GLM-5.1, an open-weight coding model built to run autonomously for hours through thousands of tool calls and hundreds of iterations. Think refactoring an entire codebase or running overnight migrations without human handholding. The model beat its predecessor on SWE-Bench Pro (58.4 vs 55.1) and ships under MIT license, meaning companies can host it internally. That's the real shift: you can now run a competitive coding AI on your own servers instead of paying per API call. The catch? You'll need serious guardrails. Long-running autonomous agents sound great until one rewrites your authentication logic at 3am.
What happened
Z.ai released GLM-5.1, an open-weight coding model built to run autonomously for hours through thousands of tool calls and hundreds of iterations. Think refactoring an entire codebase or running overnight migrations without human handholding. The model beat its predecessor on SWE-Bench Pro (58.4 vs 55.1) and ships under MIT license, meaning companies can host it internally. That's the real shift: you can now run a competitive coding AI on your own servers instead of paying per API call. The catch? You'll need serious guardrails.
Why it matters
Long-running autonomous agents sound great until one rewrites your authentication logic at 3am.