Breaking

Z.ai Releases Open-Weight GLM-5.2, a 744B MoE Built for Coding

July 7, 2026 at 21:46 EDT

Foundation Models
Open Source
Software Dev & Coding

China's Z.ai (formerly Zhipu AI) released its latest large language model, GLM-5.2, as open weights under the MIT license in mid-June 2026. Built on a 744B-parameter Mixture-of-Experts (MoE) design with a context window of up to one million tokens, it emphasizes strong performance on coding and agentic tasks.

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

GLM-5.2: An Open-Weights Model Closing In on the Top Closed Labs

A ~744B-parameter MoE model under a "Pure Open" MIT license — built for long-horizon coding and agent work, posting results near GPT-5.5 and Claude Opus 4.8 at roughly one-sixth the cost.

744B

Total parameters (MoE, ~40B active)

Token context length

~1/6

Cost vs GPT-5.5 on long-horizon coding

Terminal-Bench 2.1 — score to scale

Higher is better. GLM-5.2 leaps past its predecessor and closes on the top closed model.

63.5

GLM-5.1

81.0

GLM-5.2

~85.0

Claude Opus 4.8

SWE-bench Pro — GLM-5.2 vs GLM-5.1

GLM-5.262.1

GLM-5.158.4

Strengths

MIT "Pure Open" license — no regional or commercial limits
Matches or beats GPT-5.5 / Claude Opus on long coding tasks
Outperformed Claude Code on security tasks like IDOR detection

Practical Hurdles

Full BF16 model needs tens of GB of memory — hard on consumer hardware
Heavy thinking-log output and long processing times
Max thinking-effort settings recommended for best results

The bigger picture

A "step change" for open models sharpens the open-vs-closed divide.

Developers debate which lab could next surpass GLM-5.2 to become the open-source leader — while frustration grows that some closed models champion open source in name yet stay closed in practice.

Continue reading

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