VerdictPal
Join

Models / MiniMax

MiniMax M3

MiniMax · Released 2026-05-31

MiniMax's May 2026 open-weights release. Terminal-Bench 2.0 66.0% — the #10 score globally and the strongest non-frontier-tier result on the public ledger. 1M context, $0.40 / $2.00 per 1M tokens.

Default open-weights model for self-hosted agent infrastructure. Cheaper than DeepSeek V4 Pro Max on output ($2 vs $1.60 — well, slightly higher), with comparable Terminal-Bench scores.

#15 of 51 on AA Index · current snapshot
44 AA Index · 17 benchmark rows · 8 sourcesPosition is benchmark-specific — not a cross-family or cross-source ranking.
Composite62
Context1M
Input / 1M$0.40
Output / 1M$2.0
Knowledge cutoff2026-03-01
Output speed97 t/s
TTFT (default API)1.26s
Statussolid
KnowledgeReasoningCodingAgenticLong contextTool useHuman preference
Knowledge
85
Reasoning
93
Coding
78
Agentic
66
Long context
74
Tool use
76
Human preference
75

78average across 7 tested families

Editor's note
MiniMax M3 is the dark horse of the open-weights tier — Terminal-Bench 2.0 66.0% puts it ahead of several proprietary models. The 1M context is the real differentiator for long-document work.
Benchmark placements

Where MiniMax M3 places on each public benchmark source that publishes a row. Rank counts every model with a latest row in the same test — not a universal quality score.

BenchmarkFamilyRankScoreSourceDate
IFBenchReasoning#1/ 5783Artificial Analysis2026-07-21
AA-LCRLong context#7/ 6474Artificial Analysis2026-07-21
GPQA DiamondReasoning#7/ 6993Artificial Analysis2026-07-21
MCP AtlasTool use#10/ 1069MCP Atlas leaderboard2026-06-09
LMArena (Chatbot Arena)Human preference#12/ 1375LMArena leaderboard2026-05-25
BFCLTool use#14/ 2076Berkeley Function Calling Leaderboard V42026-04-15
SWE-bench VerifiedCoding#17/ 2378SWE-bench Verified leaderboard2026-06-09
LiveBenchReasoning#18/ 2160LiveBench leaderboard2026-04-15
LongBench v2Long context#18/ 1946LongBench v2 leaderboard2026-04-01
Artificial Analysis Intelligence IndexReasoning#21/ 6744Artificial Analysis2026-07-21
MMLUKnowledge#22/ 2285MiniMax M32026-05-31
Humanity's Last ExamReasoning#23/ 6537Artificial Analysis2026-07-21
AA output speedPerformance#26/ 69Artificial Analysis2026-07-21
Terminal-BenchAgentic#28/ 6966Terminal-Bench 2.0 leaderboard2026-05-20
AA time to first tokenPerformance#38/ 69Artificial Analysis2026-07-21
SciCodeCoding#43/ 6545Artificial Analysis2026-07-21
τ³-Bench BankingAgentic#51/ 6413Artificial Analysis2026-07-21
Family context

The three highest-scoring carded models in each capability family. Where MiniMax M3 shows up, it's highlighted.

KnowledgeMiniMax M3 · 85
  1. 1
    GPT 5.5
    91
  2. 2
    Claude Opus 4.8
    91
  3. 3
    Claude Sonnet 4.6
    91
ReasoningMiniMax M3 · 93
  1. 1
    Claude Mythos Preview
    95
  2. 2
    GPT-5.6 Sol
    94
  3. 3
    Kimi K3
    94
CodingMiniMax M3 · 78
  1. 1
    GPT 5.5
    96
  2. 2
    Claude Opus 4.8
    96
  3. 3
    Claude Opus 4.6
    96
AgenticMiniMax M3 · 66
  1. 1
    GPT 5.5
    100
  2. 2
    Claude Opus 4.8
    100
  3. 3
    Claude Fable 5
    100
Long contextMiniMax M3 · 74
  1. 1
    Kimi K3
    75
  2. 2
    GPT-5.1
    75
  3. 3
    GPT-5.6 Sol
    74
Tool useMiniMax M3 · 76
  1. 1
    Claude Sonnet 4.6
    89
  2. 2
    GPT 5.5
    87
  3. 3
    Claude Opus 4.8
    87
Human preferenceMiniMax M3 · 75
  1. 1
    GPT 5.5
    95
  2. 2
    Claude Opus 4.8
    92
  3. 3
    Gemini 3.5 Flash
    91
Newest receipts

Benchmark rows added to the public ledger for MiniMax M3 in the last 120 days. Older rows live in the full table below.

17 recent
BenchmarkFamilyScoreSourceDays ago
AA-LCRLong context74%Artificial Analysis2
AA output speedPerformance97 t/sArtificial Analysis2
AA time to first tokenPerformance1.26sArtificial Analysis2
Artificial Analysis Intelligence IndexReasoning44.4%Artificial Analysis2
GPQA DiamondReasoning92.9%Artificial Analysis2
Humanity's Last ExamReasoning37.1%Artificial Analysis2
IFBenchReasoning82.86%Artificial Analysis2
SciCodeCoding45.4%Artificial Analysis2
τ³-Bench BankingAgentic12.99%Artificial Analysis2
SWE-bench VerifiedCoding78.4%SWE-bench Verified leaderboard44
MCP AtlasTool use68.9%MCP Atlas leaderboard44
MMLUKnowledge84.8%MiniMax M353
LMArena (Chatbot Arena)Human preference1376LMArena leaderboard59
Terminal-BenchAgentic66%Terminal-Bench 2.0 leaderboard64
BFCLTool use76.2%Berkeley Function Calling Leaderboard V499
LiveBenchReasoning59.6%LiveBench leaderboard99
LongBench v2Long context45.6%LongBench v2 leaderboard113
Family coverage

A high composite that hides a weak family is a trap. These bars surface the families where this model hasn't been publicly tested, and where it leads.

Knowledge85
Reasoning93
Coding78
Agentic66
Long context74
Tool use76
Human preference75
Price vs. performance
0255075100$0.10$0.30$1$3$10$30input USD / 1M tokens · log scale →← intelligence

Hover or tab any dot for name, score, and input price.

AnthropicOpenAIMistralAlibabaDeepSeekGoogleMoonshotMiniMaxxAIMetaZhipuXiaomiMeituanCohere

Full benchmark ledger

Every catalog benchmark for MiniMax M3. Scores link to the original source; gaps mean no public row exists yet.

17 sourced rows

17 of 32 catalog benchmarks have a sourced row for MiniMax M3.

Knowledge

Reasoning

Coding

Agentic

Long context

Tool use

Performance

Human preference

Changelog
  1. Released MiniMax M3 — Terminal-Bench 2.0 66.0%, 1M context.
  2. MiniMax M2.7 retired as the flagship open-weights model.
Sources