GLM-5.2 leads the open-weight leaderboards in July 2026

GLM-5.2 from Z.ai (Zhipu) was released on 16 June 2026 and is ranked the best open-source LLM of July 2026 on published leaderboards.

Reported numbers

• 744-billion-parameter mixture-of-experts architecture
• 91.2% on GPQA Diamond (graduate-level science questions)
• 62.1% on SWE-bench Pro (real software-engineering issues)

Those are the figures as published by the leaderboard we link below; we have not run these evaluations ourselves.

Reading those two benchmarks

GPQA Diamond and SWE-bench Pro test very different things, and a model that is strong on both is genuinely useful rather than merely well-tuned.

• GPQA Diamond is a knowledge-and-reasoning test written to be hard for non-experts even with a search engine. High scores indicate reasoning depth in technical domains.
• SWE-bench Pro measures whether a model can resolve real issues in real repositories — read the codebase, make a patch, pass the tests. It punishes models that can talk about code but cannot navigate a project.

The open-weight field around it

GLM-5.2 sits in a crowded top tier alongside Kimi K3 from Moonshot, DeepSeek V4 and Qwen3.6 from Alibaba. The rough split reported across trackers: GLM-5.2 leads on raw capability, DeepSeek V4 leads on price, and Qwen3.6 leads on licence freedom.

That is a genuinely good position for anyone building: three different "best" answers depending on whether your constraint is quality, cost or legal freedom to ship.