Z.ai GLM-5
Z.ai (Zhipu) launched GLM-5 as a 744B-parameter Mixture-of-Experts model (40B active), scaled up from GLM-4.5 and trained on 28.5T tokens with sparse-attention techniques to cut deployment cost. It targets complex systems engineering and long-horizon agentic work, and Z.ai claims best-in-class performance among open-source models on reasoning, coding, and agentic tasks, narrowing the gap to frontier models like Claude Opus 4.5. The weights are open-sourced on Hugging Face and ModelScope under the permissive MIT License.
GLM-5 is real and openly released: weights are public under MIT, it is served on api.z.ai and BigModel.cn, and it is usable in coding agents such as Claude Code. Z.ai's published benchmarks show it leading or near the top of open-source models on coding (SWE-bench), agentic, and reasoning suites, while still trailing the closed frontier on the hardest reasoning tests.
GLM-5 is one of the strongest pieces of evidence in 2026 that the open-weight gap to the frontier is narrowing, and the MIT license makes it genuinely usable commercially without negotiation. The benchmarks are self-reported, so treat the 'best open-source model in the world' framing as a vendor claim, but the scores are in a plausible, competitive range and the model is independently downloadable for anyone to test. For executives, the strategic point is that a Chinese lab is now shipping a permissively licensed, frontier-adjacent model that can be self-hosted, which reframes both cost and data-control conversations. The honest caveat: open weights do not erase governance, security-review, and provenance questions, and 'narrowing the gap to Claude Opus 4.5' still means it trails the leaders on the hardest tasks.
An MIT-licensed, frontier-adjacent model you can self-host changes the build-vs-buy and data-residency calculus for any organization wary of sending data to closed APIs.