Logo
MiniMax · M3 · Jun 2026

MiniMax M3 — multimodal coding model with million-token context

  • Million-token context
  • Text + image + video
  • Frontier coding focus
  • Open weights to compare
MiniMax M3

What is MiniMax M3?

MiniMax M3 is MiniMax’s open-weight flagship from June 2026. Official materials stress frontier coding, million-token context, and native multimodal together. Versus MiniMax M2.7, long-context efficiency and coding benches rise sharply. Pick MiniMax M3 in iMini Agent to use it.

Vendor
MiniMax
Released
Jun 2026
Context
1M tokens
Max output
Per platform limit
Deep thinking
Coding & multimodal focus
Best as
MiniMax-stack flagship workhorse

What’s new vs MiniMax M2.7

Long-context efficiency, coding ceiling, and multimodal—the changes to check before you make M3 your workhorse.

More affordable million-token context

Sparse-attention design makes long-material reasoning more affordable—for large repos and long specs.

Stronger coding benches

Official SWE-Bench Pro and peers show higher completion—fit as an engineering workhorse.

Native multimodal input

Text, image, and video join the same job—for demos and UI materials mixed together.

Open weights to compare

Open weights for research and private deploy evaluation; you can still run it on iMini online.

Official benchmarks

Aggregate tables and efficiency charts from the MiniMax-M3 model card—coding, collaborative Agents, GUI, multimodal, and long-context efficiency.

MiniMax M3 coding and collaboration board (upper half)
Upper coding and collaboration board. SWE-Bench Verified 80.5%, SWE-Bench Pro 59.0%, Terminal Bench 2.1 66.0%, BrowseComp 83.5%—shown with Claude Opus, GPT-5.5, Gemini 3.1 Pro, and China peers.
MiniMax M3 board midsection: tools and GUI
Mid board. SpreadSheetBench, MCP Atlas, OSWorld-Verified, and peers open office tools and computer use—beyond “can write code,” whether M3 finishes the job.
MiniMax M3 board lower half: multimodal and reasoning, plus methods
Lower board and methods. Multimodal and contest-reasoning scores with harness and timeout notes—read scores with the method; don’t hard-compare across vendors blindly.
MiniMax M3: MSA long-context efficiency vs GQA
Long-context efficiency. Vs GQA, MSA cuts Attention FLOPs ~28.4× at 1M, prefilling ~14.2× faster, decoding ~7.6× faster—whether a million-token window is usable also hinges on cost and latency.

Three common workflows

MiniMax M3 fits flagship load on the MiniMax stack.

Frontier coding

Multi-file features, hard fixes, and test loops.

Long-material reasoning

Million-token specs and multi-source document synthesis.

Video and image assistance

Advance jobs from screen recordings or screenshots.

Vs Kimi K3 and Qwen3.7 Max

All three are China-available flagships. Gaps are modality, open-weight path, and product stack.

DimensionMiniMax M3Kimi K3Qwen3.7 Max
Context1M tokens1M tokens1M tokens
ReasoningCoding & multimodal focusDeep · tunableHybrid thinking · peak
Coding / AgentsMiniMax coding flagshipKimi long-horizon coding flagshipQwen peak text
Long tool runsLong-context multimodalVision + long threadsTools & structured output
Speed postureCompletion quality firstLower tier for speedPeak quality first
Prefer whenTied to MiniMaxTied to Kimi flagshipTied to Qwen peak text

Using MiniMax M3 on iMini

Free quota, model identity, key specs, and how it differs from Kimi and Qwen.

More models

More flagship models by vendor

GPT-5.6 SolOpenAI’s depth flagship—when the team is on the GPT stack and needs maximum depth.Claude Opus 5Anthropic’s Opus flagship—for long-horizon coding and multi-step Agents on the Claude stack.Gemini 3.1 ProGoogle’s Gemini Pro tier—for multimodal, long-context work on the Google stack.Grok 4.5xAI’s Grok flagship—when you need Grok reasoning or the xAI product stack.DeepSeek V4 ProDeepSeek’s coding and reasoning flagship—a strong open-weight pick for code and hard reasoning.Kimi K3Moonshot’s Kimi flagship—for CN/EN long context and China-market delivery.GLM 5.2Zhipu’s GLM flagship—for bilingual Agents and Zhipu-stack delivery.Qwen3.7 MaxQwen Max—when you need peak Qwen text reasoning.Llama 4 MaverickMeta’s Llama 4 flagship option—when you need the Llama weight stack.Seed 2.0 LiteByteDance Seed’s lighter tier—for high-turnaround, shorter sessions.
All models

Use MiniMax M3 free on iMini Agent

Million-token context and multimodal input—for a coding flagship workhorse.

Open MiniMax M3

No MiniMax API key required.