Breaking

Z.ai Releases GLM-5.2, a 744B MIT-Licensed Model With 1M-Token Context

June 22, 2026 at 08:11 EDT

Foundation Models
Open Source
Software Dev & Coding

China's Z.ai (formerly Zhipu AI) released its latest flagship large language model, GLM-5.2, in June 2026. The 744B-parameter MoE model is offered as MIT-licensed open weights and posts performance that closes in on Western closed-source models in coding and autonomous agent tasks.

June 16, 2026 · Z.ai (formerly Zhipu AI)

GLM-5.2: Open Weights Close In on the Closed Frontier

An MIT-licensed flagship built for long-horizon coding — 1M-token context, repo-wide refactoring, and agentic bug fixes — rivaling GPT-5.5 and Claude Opus 4.8 at roughly half the cost per task.

753B

Total MoE parameters — only ~40B active per token

Token context window — 5× larger than GLM-5.1 (~200K)

2.9×

Fewer FLOPs at 1M context via IndexShare sparse attention

Cost per task — about half of GPT-5.5

Weighted cost per completed coding task

$0.42

GLM-5.2

$0.83

GPT-5.5

Long-horizon coding benchmarks

Higher is better · GLM-5.2 vs. closed frontier leaders

SWE-bench Pro

GLM-5.2 · 62.1

GPT-5.5 · 58.6

Opus 4.8 · 69.2

Terminal-Bench 2.1

GLM-5.2 · 81.0

GPT-5.5 · 84.0

Opus 4.8 · 85.0

FrontierSWE

GLM-5.2 · 74.4

GPT-5.5 · 72.6

Opus 4.8 · 75.1

What developers praise

First open-weight model usable as a "daily driver"
Grasps an entire repository for long-context coding
Strong planning, bug fixes and feature implementation
MIT license, no regional access restrictions

Open questions

No image-input support
Complaints about inference speed and rate limits
Does not fully surpass Claude Opus 4.8

Continue reading

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