Z.AI releases open-weight GLM-5.1 754B model for 8-hour agent tasks
TL;DR
Z.AI released GLM-5.1 on April 8, 2026: a 754B-parameter open-weight model under MIT license, scoring 58.4 on SWE-Bench Pro (SOTA) with an 8-hour autonomous Linux-build demo.
What changed
Z.AI released GLM-5.1 on April 8, 2026 with 754 billion parameters under MIT license. Benchmarks: SWE-Bench Pro 58.4 (SOTA), strong Terminal-Bench 2.0 results, NL2Repo leadership over GLM-5, and improved cybersecurity-suite agentic behaviour. SWE-Bench Pro was run under OpenHands with 200K context, temperature=1, top_p=0.95. GLM-5 was deprecated April 20. The headline demo wrapped GLM-5.1 in a self-review harness for an 8-hour continuous Linux-desktop build.
Why it matters
MIT licensing on a 754B-class model removes commercial-use uncertainty entirely: modification and redistribution are permitted. The agentic-first design targets long-horizon tool-use sequences rather than chat, with the self-review loop framed as the architectural distinction from GLM-5. SWE-Bench Pro SOTA at 58.4 sets the open-frontier reference point, and the reproducibility settings allow you to validate the numbers against your own harness.
What to watch for
Quota economics on the hosted API: GLM-5.1 consumes 3x during peak (14:00-18:00 UTC+8) and 2x off-peak, with a promotional 1x off-peak through end of April. Update your model name to GLM-5.1 in Claude Code or your harness of choice. Compare Terminal-Bench 2.0 head-to-head against Claude Code 2.1.69 on your real workloads before routing production traffic. Some NIM enterprise integrations have had transition friction worth verifying.
Who this matters for
- Developers: MIT license plus 754B parameters plus SWE-Bench Pro SOTA at 58.4 plus 200K context in benchmark config. The self-review harness pattern across hundreds of tool-call rounds is worth replicating in your own agent frameworks.
Harsh’s take
The 8-hour autonomous-execution demo is the part most coverage will underplay. Getting a model to stay productive over hundreds of tool-call rounds without degrading into loops, hallucinations, or spinning its wheels is an open problem in agent design. If Z.AI's 8-hour Linux-desktop build actually produces a working result, the self-review loop is a real capability and not a demo trick. The implication for long-horizon agent harnesses is significant: cap-out points move from 30 minutes to multi-hour runs.
MIT license plus 754B parameters plus SWE-Bench Pro SOTA is the cleanest open-frontier release the Chinese ecosystem has produced. Watch whether this sticks through the next two releases or whether Z.AI follows the closing path Alibaba just took with Qwen 3.6-Max. Terminal-Bench 2.0 performance against Claude Code 2.1.69 is the head-to-head that determines whether routing agent workloads to GLM-5.1 over Claude Opus 4.7 becomes economic rather than just technically interesting.
About Z.ai
View the full Z.ai page →All Z.ai updatesGo deeper
More AI news
- Daily RoundupKimi K3.1 Pro closes on Opus 5.5, Seedance 2.5 video, and agent tools you can run today
Open-weight challengers close on the frontier, ByteDance and Kling flood Replicate with video models, and agent plumbing gains budgets, keys and permissions.
- Daily RoundupUber puts 30,000 on ChatGPT Work, Canva's first image model, and the agent login fight
Enterprise agent rollouts go company-wide while banks block agents at the door, and NVIDIA floods the day with infrastructure and open model tooling.
- Daily RoundupGoogle Cloud hosts GPT-6, NVIDIA's Berlin agent stack, and a laptop coding model
Google Cloud becomes the first cloud to sell Gemini, Claude, and GPT-6 together, while NVIDIA floods Berlin with agent infrastructure and Recursive ships a laptop coding model.