MiniMax M3

Released 2026-06 · reasoning · 1.0M tokens · 4 benchmarks · Open weight

Editorial notes

Release 1 junio 2026. Nuevo flagship MiniMax. Arquitectura MiniMax Sparse Attention (MSA): GQA + block-level sparse selection. A 1M context: >9.7x prefill, >15.6x decode vs gen previa. Multimodal nativo (text+image+video in, text out), computer use. Weights ya publicos en huggingface.co/MiniMaxAI/MiniMax-M3 (~428B total / ~23B activos) bajo MiniMax Community License (custom, comercialmente restringida: atribucion + gate de autorizacion si el producto supera $20M/anio — NO es MIT/Apache). Pricing $0.30/$1.20 per MTok (≤512K, descuento 50% permanente; lista $0.60/$2.40). La tabla oficial de MiniMax (solo imagen) es agentica/coding y NO incluye GPQA-D, SWE-Verified, AIME, MMLU-Pro, LiveCodeBench ni HLE. Scores vendor (scaffolding propio, verbosity ~91M tokens): SWE-bench-Pro 59.0, Terminal-Bench 2.1 66.0, BrowseComp 83.5 (supera Opus 4.7 79.3), OSWorld 70.1. Tercero (Artificial Analysis, harness independiente): GPQA-Diamond 92.9, HLE 37 (sin tools), AA Intelligence Index 44 (set v4 vivo; el 55 era el numero de lanzamiento, pre-medicion final). Por eso solo cuenta con 2 benchmarks vendor del core-set (ambos coding/agentic) y no alcanza el gate de elegibilidad — los reasoning GPQA-D/HLE existen pero son de tercero.

Spec sheet

Company: MiniMax
Country: CN
Type: reasoning
Release: 2026-06
Context: 1.0M tokens
Params total: 428B
Params active (MoE): 23B
License: minimax-community
Pricing (openrouter): $0.3/$1.2/M
Slug: minimax-m3

Benchmarks (4)

Coding 2

Agentic 2

Independently measured (third-party) AA 2

Measured by Artificial Analysis on its own standardized harness — the vendor did not publish these core-set benchmarks. Used to compute the Frontier Index so the model can rank, but kept out of the vendor percentile pool. See Methodology.

Cite this model

BibTeX · APA

BibTeX

@misc{frontier-minimax-m3,
  title  = {MiniMax M3},
  author = {{MiniMax}},
  year   = {2026},
  note   = {Frontier Benchmarks AI atlas. Accessed 2026-06-17},
  url    = {https://frontierbenchmarks.com/models/minimax-m3}
}

APA

MiniMax (2026). MiniMax M3 [Large language model]. Frontier Benchmarks AI. Retrieved 2026-06-17, from https://frontierbenchmarks.com/models/minimax-m3

Citation reflects the atlas page, not the original model paper. For the paper, see the "Resources" section above.

Battle vs another model ← All models More from MiniMax Methodology