BREAKING
MiniMax M3 Hits #18 on Agent Arena
0
#
overall rank
0
#
open models
0
#
M2.7 prior rank
How Agent Arena Scores Models
1
Task success
↓
2
Steerability
↓
3
Error recovery
↓
4
Praise vs complaint
↓
5
Tool hallucination
M3 Benchmark Scores
SWE-Bench Pro
59
Terminal-Bench 2.1
66
KernelBench Hard
28.8
BrowseComp
83.5
Open Weights, Frontier Goals
MiniMax M3
open-weight
●
1M-token context
●
MiniMax Sparse Attention
●
~$0.30 / $1.20 per 1M
BrowseComp edge
vs closed
●
M3 scores 83.5
●
Claude Opus 4.7 at 79.3
●
Local deploy and fine-tune
Open Models Close the Gap
AI NEWS BLITZ
MiniMax M3 lands at number eighteen overall on the new Agent Arena leaderboard.