Breaking

Z.ai Releases Open Model GLM-5.2 With 1M-Token Context

June 23, 2026 at 02:40 EDT

Foundation Models
Open Source
Software Dev & Coding

Beijing-based Z.ai (formerly Zhipu AI) released GLM-5.2, a flagship large language model tuned for agentic coding and long-horizon tasks, as MIT-licensed open weights around June 13–14, 2026. The weights are distributed on Hugging Face and ModelScope, with API access available through its own z.ai and BigModel.cn platforms as well as Cloudflare Workers AI and Together.ai.

June 2026 · Z.ai

GLM-5.2: An Open-Weight Model That Passes as a Daily Driver

Beijing's Z.ai ships a 744B Mixture-of-Experts model under the MIT license — frontier-class agentic coding, a 1M-token context, and pricing a fraction of Claude or GPT. Trained on domestic Huawei Ascend chips.

744B

Total parameters
(~40B active per token)

Lossless usable context
(~5× GLM-5.1)

MIT

Open-weight license
(self-hostable)

Terminal-Bench 2.1 — climbing toward the top tier

A big jump over its predecessor; a narrow gap to the leader remains.

62.0

GLM-5.1

81.0

GLM-5.2

85.0

Claude Opus 4.8

SWE-bench Pro

62.1 vs 58.4 (GLM-5.1)

Design & front-end Arenas

Surpasses GPT-5.5 in some cases

A fraction of the cost

API pricing per million tokens — a draw versus comparable frontier models.

Input

$1.20–1.40

per million tokens

Output

$4.10–4.40

per million tokens

In some cases a fraction — down to a tenth — of comparable Claude or GPT pricing.

Why it matters

First open model rated a viable daily driver by ex-Meta, DeepMind and Microsoft leads
Builds full websites, 3D games, OS-like apps, audits and long refactors
Self-hostable, frontier-class and open-weight under MIT

Caveats

744B MoE — even quantized, single-machine local runs are hard
Text-only in many use cases (no vision)
China-based API raises latency / regulatory concerns
Long-horizon real-world maturity still under evaluation

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.