MiniMax: every model, side by side

MiniMax M3 takes text, image and video, returns text, over a 1M token context, for long-horizon agent work, coding and tool use. MiniMax Sparse Attention swaps full attention for KV-block selection, cutting per-token compute at long context to roughly a twentieth of the last generation.

MiniMax M2.7 runs multi-agent collaboration internally, planning, executing and refining across changing environments. Handles live debugging, root cause analysis, financial modelling and document generation across Word, Excel and PowerPoint. Scores 56.2% on SWE-Pro, 57.0% on Terminal Bench 2, 1495 ELO on GDPval-AA.

MiniMax M2.5 extends M2.1's coding into office work: it generates and operates Word, Excel and PowerPoint files, switches software environments, and works alongside other agents. Scores 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp, on fewer tokens than earlier generations.

A dialogue-first model for roleplay, character chat and long multi-turn conversation. Holds tone and personality, supports rich message roles, and can learn pacing from example dialogue.

MiniMax's 230B mixture-of-experts model with open weights under MIT, tuned for coding and agent work. Available from MiniMax's own platform, Hugging Face and OpenRouter.

MiniMax M2: 10B activated parameters for coding agents, compile-run-fix loops and long-horizon reasoning. Good SWE-Bench and Terminal-Bench results, and it holds reasoning continuity across multi-step tasks.

MiniMax M1, an open-weight reasoning model with 456B total parameters and 45.9B active per token. It pairs a mixture-of-experts backbone with a lightning attention mechanism to hold context up to 1 million tokens without the usual FLOP cost, and is tuned for multi-step reasoning.

MiniMax's first image generation model with character reference support for consistent character generation across multiple images. Produces high-quality, detailed images with strong compositional abilities.

High-definition text-to-speech with studio-quality output and fine-grained emotion control.