MiniMax M3

Architecture overview:
• MSA (MiniMax Sparse Attention): new sparse attention architecture with block-level KV partitioning and KV-outer-gather-Q operator design
• At 1M context: per-token compute is 1/20 of previous generation; 9x prefilling and 15x decoding speedups vs full attention
• 4x faster than Flash-Sparse-Attention and flash-moba at M3's head configuration
• Native multimodal from training step 0 — image and video input with interleaved data scaling to 100 trillion tokens
• Toggleable thinking mode: on for complex reasoning, off for fast latency-sensitive responses
• Computer use capability for desktop automation across applications and file systems
• Open-weight; weights to be released approximately 10 days after launch

Training methodology:
• Native multimodal pretraining with interleaved text and image/video sequences from step 0
• Interactive user simulator framework for coding training: simulates real developer collaboration patterns including requirement elaboration, feedback correction, and multi-turn task iteration
• Scales coding training data to scenarios beyond single-turn task execution

Performance characteristics:
• Coding agents: 59.0% SWE-Bench Pro, 66.0% Terminal-Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard
• Agentic: 74.2% MCP Atlas, top score on Claw-Eval
• Multimodal: above Gemini 3.1 Pro on OmniDocBench
• Autonomous execution: 9.4x FP8 GEMM kernel speedup over 24 hours, 1,959 tool calls (MiniMax-reported)
• PostTrainBench: 0.37 for autonomous model training across 5 benchmarks

‍

Together AI API access:
• Access MiniMax M3 via Together AI APIs using the designated endpoint
• Authenticate using your Together AI API key in request headers
• Toggle thinking mode at request time — no separate model version required
• Supports image and video input alongside text for multimodal requests
• Available on Together AI on serverless and dedicated infrastructure

‍

Coding & software engineering:
• Complex multi-file software engineering with 59.0% SWE-Bench Pro
• Long-horizon terminal execution with 66.0% Terminal-Bench 2.1
• CUDA kernel optimization, performance engineering, and low-level systems work
• Interactive multi-turn collaboration with context retained across the full session

Agentic & research workflows:
• Autonomous research tasks: paper reproduction, literature synthesis, and experiment design
• Tool orchestration with 74.2% MCP Atlas and top Claw-Eval performance
• Autonomous model training and evaluation pipelines without human intervention
• Long-horizon tasks sustained over multi-hour sessions using 1M context

Multimodal & computer use:
• Image and video input for document understanding, visual reasoning, and multimodal analysis
• Computer use for desktop automation across applications, files, and systems
• Visual coding: generating and editing interfaces from screenshots and visual references

Reasoning & long context:
• Toggleable thinking for complex reasoning tasks requiring deep analysis
• 1M token context for processing full codebases, research papers, and extended agent sessions
• Cross-document synthesis and long-context comprehension via MSA architecture

‍