Breaking

MiniMax M3 Lands 18th Overall, 5th Among Open Models on Agent Arena

June 16, 2026 at 11:35 EDT

Foundation Models
AI Agents
Open Source

China's MiniMax has placed its open-weight model MiniMax M3 at No. 18 overall and No. 5 among open models on Arena.ai's new Agent Arena leaderboard. The model climbed from No. 22 held by its predecessor M2.7, with gains credited to confirmed task success and error recovery.

May–June 2026 · MiniMax

MiniMax M3 puts open weights on the agentic map

An open-weight model from a Shanghai lab climbs to #18 overall and #5 among open models on Agent Arena — pairing frontier-level coding with a 1M-token context window and native multimodality.

#18

overall on Agent Arena (up from #22)

among open-weight models

token context window

9.4×

CUDA kernel speedup, 147 autonomous iterations

Benchmark scores

Reported across software engineering, terminal, and browsing tasks (%).

83.5

BrowseComp

66.0

Terminal-Bench 2.1

59.0

SWE-Bench Pro

28.8

KernelBench Hard

M3 edges out a closed frontier model on browsing

BrowseComp score — higher is better.

83.5

MiniMax M3
(open weight)

79.3

Claude Opus 4.7
(closed)

What Agent Arena measures

Confirmed task success • Steerability • Error recovery • Praise vs. complaint • Tool hallucination

Scored via causal inference over millions of real user sessions — not a static problem set.

Early strengths

Reliable on messy, long-horizon engineering — fixing production issues, end-to-end app building, kernel optimization. Low cost (~$0.30/$1.20 per 1M tokens) and long context make some prefer it over closed models.

Caveats

Weaker in open-weight multi-agent setups; some reviews suggest it suits experimentation more than heavy production loads. Deep user reports remain limited given the recent launch.

Continue reading

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