Z.ai drops open-source GLM-5.1 coding agent

GLM-5.1 is Z.ai's new flagship open-source agentic coding model, released under the permissive MIT license. Instead of just answering code completions, it is built to operate autonomously as a software engineer for up to eight hours on a single task.

On the SWE-Bench Pro benchmark, GLM-5.1 clocked an impressive 58.4, edging out GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3. The model weights are already available on Hugging Face for anyone to download and use commercially.

What it does

Most coding assistants hit a hard ceiling after a handful of turns. Give them more context or extra tool calls, and they usually hallucinate, loop infinitely, or abandon the original goal. GLM-5.1 is engineered to maintain objective alignment over hundreds of iterations and thousands of tool calls.

Z.ai leader Lou noted on X that while coding agents tapped out around 20 continuous steps late last year, GLM-5.1 can now handle 1,700 tool-use steps. He considers autonomous run time to be the next major curve after classic compute scaling laws.

In real tests, this endurance turns into genuine engineering output. During a VectorDBBench benchmark, the model was handed an empty Rust skeleton and instructed to optimize a vector database engine by editing code, compiling, benchmarking, and profiling over live runs.

Where Claude Opus 4.6 plateaued at 3,547 queries per second, GLM-5.1 ran through 655 iterations and over 6,000 tool calls to hit 21,500 QPS. It didn't just tweak constants: it autonomously rewrote scan logic to IVF clusters, compressed vectors, and added two-stage reranking pipelines.

When recall dipped below 95%, it diagnosed its own regression and adjusted parameters to recover accuracy. In KernelBench Level 3, it tackled end-to-end GPU kernel optimization for models like MobileNet, VGG, and Mamba inside isolated H100 Docker containers with up to 1,200 turns per task.

Specs and availability

GLM-5.1 is a massive 754-billion parameter Mixture-of-Experts model featuring a 202,752 token context window.

Because it ships under an MIT license, organizations can inspect, fine-tune, and host it internally without proprietary lock-in. You can pull the model right now from Hugging Face.

That said, running a 754B MoE architecture is no trivial task. You will need serious data center hardware, meaning multi-GPU enterprise nodes, to run it locally.

Why it matters

Frontier agentic coding has spent the last year locked entirely behind closed APIs. GLM-5.1 proves that the open-source ecosystem can not only match but occasionally beat the likes of OpenAI and Anthropic in sustained, multi-hour engineering tasks.

If your team tackles heavy database engine optimizations, automated kernel tuning, or deep legacy refactors, this is worth deploying immediately. Just don't spin up a 754B MoE cluster if all you need is quick boilerplate autocomplete.