Breaking

Z.ai Adds GLM-5.2 to Agent Arena, Plans MIT-Licensed Weights Next Week

June 13, 2026 at 09:48 EDT

Foundation Models
AI Agents
Open Source

On June 13, 2026, GLM-5.2, the latest open-source model from Z.ai (Zhipu AI), was added to Agent Arena, a platform for real-world agent evaluation. The community-driven evaluation service Arena.ai (formerly LMArena, originating at UC Berkeley) announced the listing the same day, while Z.ai officially unveiled the model's availability.

June 13, 2026 · Z.ai (Zhipu AI) · Arena.ai

GLM-5.2 Lands in Agent Arena — Open Weights Coming Next Week

Z.ai's latest open-source model joins Arena.ai's live, usage-based evaluation, ships to all GLM Coding Plan tiers today, and is set for an MIT-licensed weight release next week — built for long-horizon, agentic engineering tasks.

token context length

MIT

open-weights license

tiers: Lite / Pro / Max / Team

GLM-5.1 rank in Agent Arena

SWE-bench Verified — GLM-5 vs reference closed model

Open weights close in on a frontier closed model on agentic coding.

77.8%

GLM-5 (MIT)

80.0%

GPT-5.2 (ref.)

Predecessor profile · GLM-5

744B-param MoE (40B active) · trained on 28.5T tokens · BrowseComp 75.9% · top open entry on Vending-Bench 2 ($4,432 balance).

Watch-out

Earlier GLM versions showed "doom loops" — real-world stability depends on the harness, and mitigation is still needed beyond static benchmarks.

How Agent Arena scores models — a live workflow loop

Real user session

→

Tools: web · files · terminal

→

Complex workflow executed

→

Scored on real signals

Signals include task success, steerability, error recovery, praise-to-complaint ratio, and tool hallucination — not static test sets.

Next week: open-source weights on Hugging Face / ModelScope, plus API and pricing details. High / Max thinking-effort levels control reasoning depth — Max recommended for coding.

Continue reading

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