MiniMax iconMiniMax

Shanghai AI company with the MiniMax-M3 1M-context text model plus speech, video and music models, sold by API from $0.15 per 1M tokens

1M Token Context (M3)
$0.30 M3 per 1M Input
300M+ Users (company claim)

Overview

MiniMax is a Shanghai AI company, founded by Yan Junjie and former SenseTime colleagues, that builds text, speech, video, image and music models and sells them through its API platform. Its current flagship text model, MiniMax-M3, is a native multimodal mixture-of-experts model with about 428B total and 23B active parameters and a 1M-token context, aimed at coding and agent work; MiniMax-M3.1-Flash (220B) is the cheaper, faster option. Many of its models also ship as downloadable weights on Hugging Face (M3 under a custom "minimax-community" license, the earlier M2 under MIT), so the line between closed API provider and open-model lab is blurred. Beyond text it offers Speech 2.8, the H3 video and H3 Image models, and Music 3.0. MiniMax listed on the Hong Kong stock exchange on January 9, 2026, raising about HK$4.82 billion (about US$619 million). Also from MiniMax: Hailuo, its consumer video generator.

The Verdict

Who Should Use MiniMax?

Best For

  • Low-cost coding and agent workloads that need long context
  • Routine, repetitive implementation tasks at volume
  • Teams that want one vendor for text, speech, video and music
  • Self-hosting a large open-weight model under its community license

Not Ideal For

  • Hard debugging or architecture work (users say it can get stuck)
  • Latency-sensitive apps (users report slow responses)
  • Buyers who need a standard permissive license for M3
  • New music API projects (closed to new users since Aug 2026)

What's Great

  • Low prices after a permanent 50% cut on M3 ($0.30 in, $1.20 out per 1M)
  • 1M-token context on M3
  • Artificial Analysis put M3 at 55 on its Intelligence Index at launch, ahead of Kimi K2.6
  • Open weights for M3, M3.1-Flash and M2.x on Hugging Face

Watch Out For

  • M3 uses a custom minimax-community license, not MIT or Apache
  • Reddit users report slow output and trouble on complex debugging
  • Most benchmark figures are self-reported by MiniMax
  • Music and lyrics APIs closed to new users as of Aug 20, 2026

Pricing

View all features & details

Models

  • MiniMax-M3: about 428B total, 23B active, 1M context, native multimodal
  • MiniMax-M3.1-Flash: 220B, lower price
  • MiniMax-M2.7, M2.5, M2.1 (229B), open weights
  • Speech 2.8, H3 video (33B), H3 Image, Music 3.0

Vendor Benchmarks (M3)

  • SWE-bench Verified 80.5
  • SWE-bench Pro 59
  • MMMU-Pro 78.1
  • Self-reported by MiniMax on the model card

Company

  • Shanghai; founder Yan Junjie
  • Founded late 2021 or early 2022 (sources disagree)
  • Listed on HKEX January 9, 2026 at HK$165 a share

Real-World Usage

Adoption (company claims)

  • 300M+ individual users
  • 2M+ enterprise clients and developers
  • 230+ countries and regions

Hugging Face Downloads

  • MiniMax-M2.5: 1.13M
  • MiniMax-M2.7: 295k
  • MiniMax-M3: 142k

How It Compares

Feature MiniMax-M3 Qwen3.7-Plus Mistral Large 4
Input price per 1M $0.30 $0.40 $0.70
Output price per 1M $1.20 $1.60 $2.10
Context 1M Up to 1M 1M
Open weights Yes, custom license API model (open Qwen3.8 weights separate) Yes, modified MIT
Maker MiniMax (China) Alibaba (China) Mistral AI (France)

Sources: MiniMax · Pricing · MiniMax-M3 card · Artificial Analysis · SiliconANGLE, IPO (Jan 2026) · Qwen pricing · Mistral pricing

Is this your tool? Add a badge