Open Source · JUNE 13, 2026
Z.ai Ships GLM-5.2 With 1M-Token Context and an MIT Pledge — and No Benchmarks
Zhipu's international brand pushed its 744B MoE flagship to a million-token window and added dual thinking-effort presets, but launched without a single score and gated the weights behind a 'next week' promise.
Z.ai pushed GLM-5.2 live on June 13, 2026 with a million-token context window, dual High/Max thinking-effort presets, and not a single benchmark number to back any of it. The model ships immediately on every tier of the GLM Coding Plan, starting from roughly $18/month for Lite. The standalone API, the consumer chatbot, and the MIT-licensed weights are all promised for "next week."
That sequencing is the story. Six weeks ago, GLM-5.1 launched with a self-reported Code Arena Elo of 1530 (third worldwide, behind two Claude Opus 4.6 variants) and a 58.4 on SWE-bench Pro, edging GPT-5.4's 57.7. The numbers were the pitch. This time, Z.ai is asking developers to take the model on vibes and a license promise.
The architecture is unchanged from February's original GLM-5: a 744-billion-parameter Mixture-of-Experts, now stretched to a 1,000,000-token window (5× the 200,000 of GLM-5.1) with a maximum output of 131,072 tokens per response on the glm-5.2[1m] variant. Z.ai's developer docs are careful to call the context "usable," explicitly distinguishing comprehension across the full window from merely accepting the tokens. It's a hedge that reads as a tell: long-context claims have become the most legible-to-buyers, hardest-to-verify spec in the model market, and the company knows it.
GLM-5.2 is the fourth flagship-tier release in the GLM-5 line in roughly four months, after the original on February 11, GLM-5-Turbo on March 15, and GLM-5.1 on April 7. The release tempo is the point. So is the channel: first-day support for Claude Code, Cline, OpenCode, Roo Code, Goose, Crush, OpenClaw, and Kilo Code via an OpenAI-shaped chat completions API. Z.ai isn't selling a chatbot. It's selling the substrate underneath someone else's agent.
The competitive frame is open-weight Chinese coding models against each other and against the closed American frontier. Moonshot's Kimi K2.7-Code and Alibaba's Qwen 3.7 Max are the immediate neighbors; the MIT pledge, when the weights actually drop, is what Z.ai will use to peel developers off both. GLM-5.1's weights arrived about eleven days after its API launch, so "next week" has prior art. The free chat.z.ai endpoint, meanwhile, is still serving GLM-5.1.
Zhipu, the Tsinghua spinout behind the Z.ai brand, completed its Hong Kong IPO on January 8, 2026, raising about HKD 4.35 billion (~USD 558 million) at a market cap near USD 52.83 billion. That capitalization buys this cadence. It also raises the bar on what each release has to actually prove, which makes shipping without numbers an interesting choice. The benchmarks will arrive when the weights do, or they won't, and that gap is itself the disclosure.