Breaking

Z.ai Releases GLM-5.2, Rated On Par With Top US Models

June 28, 2026 at 19:21 EDT

Foundation Models
Open Source
Software Dev & Coding

China's Z.ai (Zhipu AI) released its flagship AI model GLM-5.2 in June 2026. Offered as open weights under the MIT license, it is rated as matching or in some cases beating leading public models from US labs, and is seen as the first Chinese model to reach that bar.

June 2026 · Z.ai (formerly Zhipu AI)

China's GLM-5.2 closes in on US frontier models — and was trained on Huawei chips

An open-weight MIT-licensed model built for hours-long agentic coding, released right after US export curbs. Reported to match — and in places beat — Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro.

744B

Total parameters (MoE) — ~40B active

Token context, stably handled

~50%

Of competitor pricing, in some cases

MIT

Open weights, deploy locally

Terminal-Bench 2.1 — the leap over GLM-5.1

Score (higher is better). GLM-5.2 jumps +17.5 points to land within reach of the closed leaders.

63.5

GLM-5.1
(prior)

81.0

GLM-5.2
(open)

~84

GPT-5.5

85.0

Opus 4.8

SWE-bench Pro — edging past GPT-5.5

GLM-5.2

62.1%

GPT-5.5

58.6%

GLM-5.1

58.4%

Why this release landed hard

TIMING

Arrived just after Anthropic restricted exports of its advanced models.

HARDWARE

Reportedly trained solely on Huawei chips — no US NVIDIA GPUs.

STRATEGY

Open weights countering US chip and model export controls.

PRAISE

An open "daily driver," called "Opus-class" on frontend
1st on Design Arena, over GPT-5.5; 2nd on FrontierSWE
Strong cost-performance; easy to customize and self-host

CONCERNS

Reasoning accuracy can lag GPT and Opus
Reports of hallucinations and logical errors
Misuse risk flagged for cybersec

Continue reading

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