Model

MiniMax M2.7

OPEN

MiniMax · MiniMax M2.7 family · released Mar 18, 2026

Agentic coding/reasoning model billed by MiniMax as its first model to autonomously drive a meaningful share of its own RL research/development workflow ("self-evolution"). Open weights on Hugging Face; also served via a 2x-priced 'M2.7-highspeed' deployment tier (same weights, faster serving) which is not modeled as a separate entry here.

ReasoningCodingVisionFunction callingTool useAgentic

2123.7

Elo · rank #67

Parameters

230B

Active params

10B (MoE)

Context

200K tokens

Architecture

Sparse Mixture-of-Experts, 256 local experts (8 active/token), 62 layers, RoPE + QK-RMSNorm attention; 230B total / 10B active params (~4.3% activation)

License

MiniMax Model Community License (non-commercial use restrictions; commercial use requires separate agreement) — verify exact license text at https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE

Languages

—

API price (in/out)

$0.3 / $1.2

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

Agents' Last ExamAgents14.2%#25

best: GPT-6 Astra · 59.3%

AIMEMath94.2%#22

best: GPT-5.2 · 100.0%

BrowseCompAgents77.8%#20

best: Kimi K3 · 91.2%

GDPval-AAAgents1495#23

best: Claude Fable 5 · 1932

GPQA DiamondReasoning87.0%#52

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning28.0%#56

best: Claude Opus 5 · 64.7%

IFBenchReasoning76.0%#19

best: MiniMax M3 · 83.0%

MMLU-ProKnowledge81.8%#48

best: Claude Fable 5 · 91.5%

SWE-bench ProCoding56.2%#30

best: Claude Fable 5.1 · 81.2%

SWE-bench VerifiedCoding75.4%#43

best: Claude Opus 5 · 96.0%

Terminal-Bench 2.0Coding55.0%#53

best: Gemini 3.8 Flash · 89.4%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Research only

QLoRA156.4 GB1× B200 192GB

LoRA489.9 GB4× H200 141GB

Full fine-tune3691.5 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $3.58 (1× B200 192GB)

Plan a fine-tune →

MiniMax M2.7 family

Elo progression across releases

MiniMax M2.7Mar 20262123.7

Resources

HFMiniMaxAI/MiniMax-M2.7 DOCShttps://platform.minimax.io/docs/release-notes/models APIAvailable via MiniMax API

API price $0.3/$1.2 · each benchmark row carries its own source badge (see methodology)