Z.ai releases GLM-5.1 — a 754B open-source LLM that it says outperforms GPT-5.4 and Claude Opus 4.6

The release

Z.ai (aka Zhupai AI) has published GLM-5.1 under a permissive MIT license on Hugging Face, making the model available for commercial download and modification. The company bills GLM-5.1 as a 754-billion-parameter Mixture-of-Experts model with a 202,752-token context window — a heavyweight by any measure. It has been reported that GLM-5.1 bested GPT-5.4 and Anthropic’s Claude Opus 4.6 on SWE-bench Pro; those performance claims are from Z.ai and independent verification is still pending.

What the company says it can do

Z.ai frames GLM-5.1 as an “agentic” model built for sustained autonomous work — allegedly able to run on a single task for up to eight hours and execute thousands of tool calls without losing alignment. The company’s lead posted that models used to manage only a few dozen steps; “glm-5.1 can do 1,700 rn,” he wrote, and the pitch is that autonomy over time, not raw speed, is the next big curve in AI. The stakes are high: Z.ai listed on the Hong Kong Stock Exchange earlier this year and is using this open release to stake a claim as a major independent LLM developer in the region.

The test claims and architecture

Z.ai’s technical write-up points to a “staircase” optimization pattern rather than steady, diminishing returns — periods of incremental tuning punctuated by structural leaps. In a featured benchmark called VectorDBBench, the company reports moving from 3,547 queries/sec to 21,500 qps after thousands of tool calls and hundreds of iterations, claiming breakthroughs like IVF cluster probing, two-stage pipelines, and quantized routing along the way. Those results are impressive on paper, but they are company-provided; independent benchmarking by the wider research community will be the true litmus test.

Why it matters

Open-source AI gets another headline-grabbing contender, and with an MIT license this one invites the community to poke, prod, and prod some more. Is the future sprinting for latency or running a marathon of sustained autonomy? Z.ai is betting on the latter — and on the idea that the open-source world will validate (or refute) its bold claims. Either way, expect more benchmarking, more forks, and lots of spirited debate.

Sources: venturebeat.com