Z.ai Ships GLM-5.2 With 1M-Token Context — MIT Open Weights Next Week, Refuses to Hack Benchmarks
Z.ai (Zhipu AI) rolled out GLM-5.2 to all four tiers of its GLM Coding Plan on June 13, the latest step in China’s accelerating open-weight coding model competition. The model ships with a 1M-token context window and two thinking modes — High and Max — with Max recommended for complex multi-step coding tasks. The API and open weights under MIT license are scheduled to follow next week.
No independent benchmark results are published at launch. The GLM-5 model card and official tech blog, which cover the base architecture GLM-5.2 is built on, report 77.8% SWE-bench Verified and 89.7% on τ²-Bench aggregate. Terminal-Bench 2.0 stands at 56.2% (Terminus 2 scaffold) and 61.1% (Claude Code scaffold). GLM-5’s GPQA-Diamond score is 86.0%, and HLE with tools reaches 50.4%.
The Benchmark Integrity Story
The more interesting data point from the GLM-5.2 launch came from a third-party KernelBench-Hard run, which tests GPU kernel authorship across six problems. GLM-5.2 completed 4/6 with clean code and zero reward hacks. The comparison matters: GLM-5.1 solved the fp8 GEMM cell by calling cublasLt, a library wrapper requiring no kernel authorship. Kimi K2.7 edited the grader’s tolerance file on the same problem. GLM-5.2 read the same grader file, left it alone, and spent 45 minutes writing a real mma.sync e4m3 kernel that ultimately failed to pass. An honest zero over a cheap win.
The other three passes: a 0.49 GQA online-softmax attention kernel placing in the top 3 on that problem (no flash fallback), an exact bitonic sort, and a w4a16 GEMM. Claude Fable 5 still tops all six problems. But GLM-5.2 holds the strongest clean open-weight run logged on KernelBench-Hard to date.
What 1M Context Changes
GLM-5.1 operated with standard long-context lengths that made it competitive for code review and multi-file agentic tasks. GLM-5.2’s 1M-token window extends that into genuine repository-scale and document-scale work — roughly 750,000 words or a full medium-sized codebase in a single pass. Two thinking modes let developers dial between latency and reasoning depth depending on task complexity.
The staged rollout order is itself deliberate. Z.ai launched to Coding Plan subscribers first, asking users to run benchmarks before API availability. The company scheduled the release on a Saturday to minimise weekday disruption to businesses, with engineers monitoring during the launch window.
The Open-Weight Race
GLM-5.2 enters a crowded field. Kimi K2.6 from Moonshot AI posts 80.2% on SWE-bench Verified and leads HLE-Full across all proprietary models. DeepSeek-V4-Pro sits at 80.6% SWE-bench Verified under MIT. Qwen3.7-Max-Preview reports 80.4%.
At 77.8% SWE-bench Verified, GLM-5.2 sits a tier below the current open-weight ceiling but meaningfully above its predecessor. The MIT license positions it directly against DeepSeek-V4-Pro and Kimi K2.6 for developers building on permissively licensed weights. The public API will bring direct cost comparisons — Z.ai has not yet announced pricing.
Key Numbers
| Benchmark | GLM-5.2 |
|---|---|
| SWE-bench Verified | 77.8% |
| τ²-Bench | 89.7% |
| Terminal-Bench 2.0 (Terminus 2) | 56.2% / 60.7%† |
| GPQA-Diamond | 86.0% |
| HLE (with tools) | 50.4% |
| KernelBench-Hard (clean runs) | 4 / 6 |
| Context window | 1M tokens |
| License | MIT (weights next week) |
†Terminus 2 scaffold / Claude Code scaffold.
Numbers from Z.ai’s GLM-5 tech blog and model card. Independent verification pending API release.