Skip to main content

MiniMax-M2.5

MiniMaxMiniMaxOpen WeightMIT · Commercial OK

Description

MiniMax M2.5 is the world's first production-level model designed natively for Agent scenarios. Building on the M2.1 foundation, M2.5 delivers significant improvements in programming, tool calling, search, and office productivity. With only 10B activation parameters from its 230B MoE architecture, it achieves competitive performance against top international models like Claude Opus 4.6 while maintaining high throughput and efficient inference. M2.5 supports full-stack development for PC, App, and cross-platform applications, and excels in agentic workflows including automated customer support, data-analysis pipelines, and complex task execution.

Release Date
2026-02-12
Parameters
230.0B
Context Length
197K
Modalities
image, text

Capability Radar

37
general
38
coding
85
reasoning
57
scienceest.
70
agents
60
multimodal

Science uses a reasoning proxy when dedicated science benchmarks are unavailable.

Rankings

Domain#RankScoreSource
Agents & Tools27
65.0
LS
Code Ranking68
68.0
AA
General Ranking46
79.0
AA
Science71
67.0
AA

Benchmark Scores (LLM Stats)

Agents

BrowseComp76.3%SR
MEWC74.4%SR
SWE-Bench Pro55.4%SR
VIBE-Pro54.2%SR

Code

SWE-Bench Verified80.2%SR
Multi-SWE-Bench51.3%SR

Finance

GDPval-MM59.0%SR

General

BFCL_v3_MultiTurn76.8%SR

AA Evaluation Indices

Intelligence Index
41.9
Coding Index
37.4
Tau2
1.0
Gpqa
0.8
Ifbench
0.7
Lcr
0.7
Scicode
0.4
Terminalbench Hard
0.3
Hle
0.2

LLM Stats Category Scores

Frontend Development
80
Search
80
Agents
70
Code
60
Finance
60
General
60
Multimodal
60
Reasoning
60

Pricing

Input Price$0.3 / 1M tokens
Output Price$1.2 / 1M tokens
Blended Price (3:1)$0.525 / 1M tokens

Speed

Tokens/sec84.0 tokens/s
Time to First Token1.44s
Time to Answer25.25s

Available Providers

(LS internal units)
ProviderInput PriceOutput Price
MiniMax300K1.2M

External Sources