Overview
- Tool name: MiniMax M3 (MiniMax-M3)
- Developer: MiniMax
- Official website: https://www.minimax.io (international) / https://platform.minimax.cn (China)
- Category: AI Models
- License: MIT (open weights)
MiniMax M3 is the company's current flagship: a natively multimodal coding model with a 1M-token context window focused on frontier coding. The previous M2.5/M2.7 generations have a published retirement schedule (shutting down November 18, 2026), with M3 as the official migration target.
MiniMax's M2 line (230B MoE, 10B active) went viral in the open-source community for near-frontier capability at very low prices; M3 continues that playbook, upgraded to multimodal.
API pricing (official, per million tokens)
- MiniMax-M3 (up to 512K input): list $0.60/$2.40, $0.30/$1.20 after a permanent 50% discount; cache reads $0.06
- MiniMax-M3 (above 512K input): list $1.20/$4.80, discounted to $0.60/$2.40
- Priority tier (faster admission): 1.5x standard
- China platform: M3 up to 512K at CNY 2.10/8.40 after discount
Pricing verified against the official minimax.io pricing page (October 2026).
Who it is for
- Coding and refactoring: the M line's specialty, with mature tooling (Claude Code-compatible workflows and more)
- Long-context multimodal: 1M-token window plus native image/video input
- High-throughput agents: half-price rates and $0.06 cache reads keep agent loops extremely cheap
- Self-hosting: MIT weights, one of China's open-source big five
Limitations
- M2.5/M2.7 retire on November 18, 2026; existing integrations need migration
- General reasoning benchmarks trail the Qwen3.5 / DeepSeek V4 flagships; the strength is coding and interaction
- Multimodal output (speech/video) lives on MiniMax's separate model lines (speech-2.8, the H3 video models)
Frequently asked questions
What is the latest MiniMax model?
MiniMax M3 (multimodal coding flagship, 1M context). M2.5/M2.7 retire on November 18, 2026; an M3.1 Flash Preview exists for subscription plans.
What does the MiniMax M3 API cost?
$0.30 input / $1.20 output per million tokens after the permanent 50% discount (up to 512K input), cache reads $0.06 (official minimax.io pricing).
Is it open source?
The M line ships open weights under MIT, free for commercial use and self-hosting.
MiniMax M3 vs Kimi K3 vs DeepSeek V4?
MiniMax M3 for the cheapest coding tier ($0.30/$1.20); Kimi K3 for a 1M-context coding flagship ($3/$15); DeepSeek V4 Flash for extreme general-purpose value ($0.14/$0.28).
*This page is not affiliated with MiniMax. Pricing verified as of October 2026 (minimax.io); confirm current figures on the official site.*
