MiniMax M3

MiniMax's frontier multimodal coding model with 1M context — positioned for agentic tool use and long-horizon coding, available via MiniMax API (Anthropic-compatible text API surface).

Announced Jun 2026 on X as first open-weights model combining coding frontier, agentic frontier, and 1M sparse attention (9,978 likes, 2,618 bookmarks). Source: X/@MiniMax_AI, 2026-06-01

Official docs list M3 as: "Frontier multimodal coding model with 1M context window — Multimodal • 1M context • Frontier coding." Source: MiniMax API Docs, https://platform.minimax.io/docs/guides/models-intro, 2026-06-13

Claimed benchmarks (launch artifact — verify independently)

Benchmark Score
SWE-Bench Pro 59.0%
Terminal Bench 2.1 66.0%
SWE-fficiency 34.8%
VIBE V2 50.1%
SVG-Bench 63.7%
KernelBench Hard 28.8%
BrowseComp 83.5%
GDPval rubrics 74.7%
BankerToolBench 76.1%
MCP Atlas 74.2%
OSWorld-verified 70.0%

The local artifact is a benchmark card comparing M3 against Opus 4.7, GPT 5.5, and Gemini 3.1 Pro. Treat the numbers as vendor launch claims until Kevin's eval loop can reproduce the relevant tasks. The durable signal is not "switch models"; it is "candidate route for long-context multimodal agents that need coding, browsing, UI, diagram, video, or MCP-heavy evaluation." Source: X/@MiniMax_AI, 2026-06-01

Version snapshot

Artifact Snapshot
MiniMaxAI/MiniMax-M3 Hugging Face SHA bfd6c97f0296da547f10ecb20102c5d51a5c462e; created 2026-06-02; modified 2026-06-23; 192,311 downloads; 1,267 likes.
MiniMaxAI/MiniMax-M3-MXFP8 Hugging Face SHA 1c4e6a69f3278df8dd6c9693fcca241efe439b7d; created 2026-06-02; modified 2026-06-15; 572,278 downloads; 43 likes.
License / runtime caveat Hugging Face tags show license:other, custom_code, minimax_m3_vl, multimodal, MoE, agent, coding, video, and arxiv:2606.13392. Use an isolated eval sandbox before loading custom code or wiring into production agents.

Source: Hugging Face model API, 2026-06-30; Source: Hugging Face MXFP8 model API, 2026-06-30

API access

  • Platform: platform.minimax.io
  • Text API docs: Anthropic-compatible endpoint (see text-anthropic-api in MiniMax docs)
  • Launch promo: 50% off ≤512K context first 7 days; priority via api@minimax.io. This promotion was time-boxed from the 2026-06-01 launch thread and should not be treated as current pricing after 2026-06-08 without rechecking MiniMax. Source: X/@MiniMax_AI thread, 2026-06-01

Stack fit (Kevin)

Use Recommendation
Cost-sensitive long-context agent runs Evaluate vs Composer 2.5 / Codex on Terminal Bench tasks
MCP-heavy harness loops MCP Atlas score is the relevant signal — benchmark on Dedalus tool suite
Production default No — treat as candidate until eval loop gates pass (The Eval Loop (Slop Is an Output Problem))
AI SDK wiring Route via Vercel AI SDK provider if/when official adapter exists

Track alongside MiniMax M3 mentions in harness comparisons; re-run /last30days MiniMax M3 before betting production workloads.


Timeline