GLM-5.2 Arrives With a Usable 1M-Token Context and MIT Open Weights

A Coding-First Open Model Stretches to a Full Million Tokens

The open-weight large language model landscape gained a notable new entry this week. On June 13, 2026, the lab Z.ai released GLM-5.2, a coding-focused model whose headline feature is a usable one-million-token context window — roughly a fivefold leap over its predecessor's ~200,000-token ceiling. For developers who have watched open models steadily close the gap with proprietary systems, GLM-5.2 is a meaningful marker on that trajectory.

What a Million-Token Context Window Unlocks

Context window is the amount of text a model can hold in working memory at once. A one-million-token window means GLM-5.2 can ingest an entire mid-sized codebase, lengthy design documents, and conversation history simultaneously, without the lossy summarization that smaller windows force. For agentic coding workflows — where a model navigates many files, traces dependencies, and edits across a project — that expanded context translates directly into fewer dropped threads and more coherent multi-step reasoning.

Inside the Architecture

GLM-5.2 is built as a Mixture-of-Experts (MoE) model with 744 billion total parameters, of which roughly 40 billion are active per token. That sparse-activation design is the prevailing approach for balancing capability against inference cost: the model carries a large knowledge capacity while only engaging a fraction of its parameters on any given token. It supports up to 131,072 output tokens and offers two "thinking effort" levels, High and Max, letting developers trade latency for deeper deliberation on hard problems.

MIT-Licensed Weights and Why the License Matters

Z.ai made GLM-5.2 available immediately across its GLM Coding Plan tiers, with MIT-licensed open weights slated to follow. The MIT license is among the most permissive in software — it allows commercial use, modification, and self-hosting with minimal restriction. For teams with data-residency requirements or a preference to run inference on their own infrastructure, openly licensed weights are the difference between a tool they can fully control and one they can only rent.

Drop-In Compatibility With Agentic Tooling

One practical detail stands out: GLM-5.2 is reachable through an Anthropic-compatible endpoint, meaning it can slot into agentic developer tools such as Claude Code, Cline, and OpenCode with little more than a base-URL and model-name swap. That interoperability lowers the switching cost dramatically and reflects a broader maturation of the open ecosystem, where models increasingly speak common API dialects.

A Measured Note on Benchmarks

In the interest of accuracy, it is worth flagging that Z.ai did not publish formal benchmark results at launch. The capability claims around GLM-5.2 will be validated as the community runs it through standard coding and reasoning evaluations in the coming weeks. As we often emphasize in our AI coverage, confirmed specifications and independent benchmarks are distinct things — and the open-weight release is precisely what makes that independent verification possible.

The Bigger Picture for Open AI

GLM-5.2 continues a genuinely encouraging trend: capable, permissively licensed models arriving with long context and real agentic utility. Whether your interest is self-hosted LLM hardware or building coding agents, the expanding menu of open options is good news for builders everywhere.

Sources: MarkTechPost (June 14, 2026); Tony Reviews Things (June 14, 2026); Codersera (June 14, 2026).