GLM-4.6

GLM-4.6 is Z.ai's open-weight flagship coding model (357B-parameter MoE, MIT license), released September 2025. It expanded the context window from 128K to 200K tokens and targets agentic coding and tool-use workloads — it ships natively in coding agents like Claude Code, Cline, Kilo Code and Roo Code, with vLLM/SGLang support for local deployment. Z.ai reports gains across agents, reasoning and coding benchmarks, near parity with Claude Sonnet 4 on its extended CC-Bench evaluation (48.6% win rate, ~15–30% fewer tokens than GLM-4.5), and competitive public scores (SWE-bench Verified ~68%, AIME-25 ~98.6% per third-party roundups). Weights are published on Hugging Face and ModelScope under MIT, making it one of the cheapest ways to run a frontier-class coding model — including self-hosted.

Provider-reported (Z.ai release materials, reproduced in third-party roundups): SWE-bench Verified 68.0% · LiveCodeBench v6 82.8% · AIME-25 98.6% · GPQA ~82.9% · CC-Bench (Z.ai's own 74-task Claude Code harness) 48.6% win-rate vs Claude Sonnet 4.

Independent: OpenRouter runs its own GPQA-Diamond / TAU-Bench evals per provider endpoint: GPQA-Diamond 76.9–84.4%, TAU-Bench 69–75.8% (varies by provider). Third-party comparison pieces (allaboutai, digitalapplied) corroborate the "slightly below Sonnet 4.5, far cheaper" positioning using public benchmark data.