Updated: August 2026 — MIT weights confirmed, benchmark results available, and a community project now lets you run GLM-5.2 on a consumer PC.
GLM-5.2 is Z.ai’s MIT-licensed open-weight coding model — 744B parameters, 1M-token context — that scores 51 on the Intelligence Index v4.1, beating DeepSeek V4 Pro and Kimi K2.6. Released June 13, 2026, it costs ~$1.40/$4.40 per million tokens via OpenRouter and runs on consumer hardware via the community Colibri project.
GLM 5.2 Release Brings 1M Context and MIT License to Coders
Z.ai (formerly Zhipu AI) shipped GLM-5.2 on June 13, 2026, rolling it out immediately to every tier of its GLM Coding Plan — Lite, Pro, Max, and Team. The release targets the coding-agent market directly, with compatibility for eight major developer tools and a context window large enough to hold entire codebases in a single prompt.
The timing is notable: GLM 5.2 arrives the same week the US government banned Anthropic’s Fable 5 for non-US users. For developers who suddenly lost access to their preferred coding model, GLM 5.2 is an immediate, globally available alternative — no export controls, MIT license pending.
744B Parameters, 1M Tokens, 131K Output
GLM 5.2 runs on the same 744-billion-parameter Mixture-of-Experts architecture as GLM-5, according to Codersera’s analysis. The MoE design activates only a fraction of those parameters per token — frontier-scale capability without frontier-scale inference cost.
The context window is labeled glm-5.2[1m] in the API: one million tokens. Output per response jumps to 131,072 tokens — roughly five times GLM-5.1’s output ceiling. That matters for long agentic tasks where the model must generate entire files, not just snippets.
Two thinking modes ship at launch: High and Max. AI Weekly reports Zhipu recommends Max for complex coding tasks, with High trading reasoning depth for speed on simpler jobs.
Eight Coding Tools That Work With It Today
GLM 5.2 is compatible out-of-the-box with Claude Code, Cline, OpenCode, Roo Code, Goose, Crush, OpenClaw, and Kilo Code — per Codersera’s verified list. That covers virtually every major AI coding assistant in active developer use.
Zhipu is deliberately targeting the developer-tools ecosystem built around Claude Code. Dropping a compatible model with a larger context window is a calculated move to capture developers who lost Fable 5 access overnight.
MIT License Confirmed: No Regional Locks, No Nvidia Required
GLM-5.2’s MIT open weights are now publicly available with no usage restrictions or regional locks. The model runs entirely on Huawei Ascend silicon — not Nvidia — which makes it immune to US chip export controls. Z.ai (the company’s current brand name after rebranding from Zhipu AI) is on the OFAC sanctions list, so US-based organisations should treat API access as a legal gray area and consult counsel before commercial deployment. Developers outside the US face no such restriction. The MIT license itself imposes no geographic limits on who downloads and runs the weights.
A Q4-quantised version of the 744B MoE model runs on high-end consumer hardware — and the community-built Colibri project (see below) goes further, enabling it on ordinary consumer PCs without enterprise-grade GPUs. Community fine-tuning is already underway on Hugging Face following the weights’ public release.
GLM 5.2 Release Has One Glaring Gap: No Benchmarks
Zhipu has not published independent benchmark results. No HumanEval. No SWE-bench. No MMLU or BigCodeBench numbers. Launch materials claim “powerful coding” and “strong long-horizon” performance — which is what every model release says.
The Aider.chat coding leaderboard and BigCodeBench would be the right places to check once community evals appear. Until then, performance claims are vendor assertions. Not a dealbreaker for developers willing to test it themselves — but not a basis for production decisions either.
Can You Run GLM-5.2 on a Consumer PC?
Yes — a community GitHub project called Colibri runs GLM-5.2 on slow consumer PCs, including ordinary Windows and Mac hardware without enterprise-grade GPUs. Colibri is not an official Z.ai product; it is an open-source community tool. The project hit a Hacker News score of 862 on July 11, 2026 — one of the top-10 stories that day — reflecting genuine developer interest in local, laptop-friendly frontier AI. For developers who want to avoid cloud API costs or who work in air-gapped environments, Colibri closes the last practical barrier to running GLM-5.2 in production. The MIT licence means there is no restriction on self-hosting.
Frequently Asked Questions
What is the GLM 5.2 release?
GLM-5.2 is a large language model released by Z.ai (formerly Zhipu AI) on June 13, 2026. Built on a 744-billion-parameter Mixture-of-Experts architecture, it features a one-million-token context window and two thinking modes (High and Max). It is designed primarily for coding tasks and is available on Zhipu’s GLM Coding Plan.
Is GLM 5.2 open source?
GLM-5.2 is available via API and as downloadable MIT-licensed open weights with no usage restrictions or regional locks. The weights are publicly available and can be run locally. A community project called Colibri makes it possible to run GLM-5.2 on ordinary consumer PCs. Note: Z.ai (formerly Zhipu AI) is on the US OFAC sanctions list — US organisations should verify compliance before commercial use.
How does GLM 5.2’s context window compare to other models?
GLM 5.2 offers a one-million-token context window, among the largest in any open-weight model. Its output cap of 131,072 tokens per response is approximately five times GLM-5.1’s output ceiling. Many frontier models cap output at 8,192 to 32,768 tokens by comparison.
Does GLM 5.2 work with Claude Code?
Yes. GLM 5.2 is compatible out-of-the-box with Claude Code, along with seven other coding tools: Cline, OpenCode, Roo Code, Goose, Crush, OpenClaw, and Kilo Code. Developers can point their existing coding agent to GLM 5.2 as the backend model without changing their workflow.
Has GLM 5.2 been independently benchmarked?
No. As of launch, Zhipu has not published independent benchmark results on standard evaluations like HumanEval, SWE-bench, or BigCodeBench. Performance claims are from the vendor only. Independent community evaluations are expected once the open weights are released.
GLM 5.2 arrives with a million-token context, compatibility with eight coding tools, and an MIT license on the way. The missing piece is independent evals. Once those land, this story gets a lot more interesting. For now, if you lost access to Anthropic’s Fable 5 this week, GLM 5.2 is the most capable open alternative available.
GLM 5.2’s MIT licence puts it in the small group of models a business can use commercially without a revenue trigger, though its 744 GB VRAM requirement rules out single-machine self-hosting. Our guide to the 25 best free AI tools that actually work compares that hardware floor against the Apache-2.0 models that do run on one GPU.
