Breaking

Z.ai Releases 744B-Class Open Model GLM-5.2 Under MIT License

June 28, 2026 at 00:13 EDT

Foundation Models
Open Source
Software Dev & Coding

China's Beijing-based AI lab Z.ai (formerly Zhipu AI) unveiled its flagship large language model "GLM-5.2" in June 2026, releasing the model weights publicly under an MIT license. The open-weight model is optimized for long-horizon coding tasks and is drawing attention for delivering performance approaching that of US closed models at low cost.

June 2026 · Zhipu AI (Z.ai)

GLM-5.2: an open-weight model closing in on the frontier

Released under the MIT license with a 1M-token context window and a focus on agentic coding, China's flagship open model is being praised by Silicon Valley developers as rivaling closed systems — at a fraction of the cost.

~750B

Total parameters (MoE), ~40B active per token

Practical lossless context length (tokens)

MIT

License — free to fine-tune & self-host

Benchmark scores vs. references

Same scale (0–100) — each column built from stacked blocks

81.0

GLM-5.2
Terminal-Bench 2.1

85.0

Claude Opus 4.8
Terminal-Bench 2.1

62.1

GLM-5.2
SWE-bench Pro

58.4

GLM-5.1
SWE-bench Pro

Top open-weight on the Artificial Analysis Intelligence Index (51) — within striking distance of closed frontier models.

Cost per million tokens

$1.40

Input tokens

$4.40

Output tokens

Substantially cheaper than closed models — and runnable locally via quantized GGUF on a Mac Studio.

What developers praise

"Found bugs other models missed"
"Matches or exceeds Opus 4.8" on long coding tasks
Strong for large-repo reviews & privacy-focused local security audits

What raises caution

Removable safety guardrails & misuse risk
Occasionally mixes in Chinese output
Benchmarks lean on official figures — real-world stability needs verification

Continue reading

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