Latest28 Sep 2026Coding model
MiniMax-M3.1-Flash-Preview
MiniMax's newest M-series text model, a preview tuned for everyday coding work from bug fixes to complete features. It runs only inside the MiniMax Code app for subscribers on MiniMax's M Plan, with no pay-as-you-go API price or open weights yet.
- 1M-token context window
- Tunable thinking depth
- Multimodal input; MiniMax calls it a multimodal coding model
- Built for bug fixes through complete feature implementation
AccessOnly inside the MiniMax Code desktop app for M Plan / Token Plan subscribers ('available only through M Plan and MiniMax Code for now'); no pay-as-you-go API price and no open weights
31 Jul 2026Video model
MiniMax H3
MiniMax's third-generation video model. It takes text, images, video and audio as input and produces video with native stereo sound, up to 15 seconds at 2K. It covers text-to-video, image-to-video, audio-to-video and editing tasks, and MiniMax pitches it on price per second of output.
- Inputs: text, image, video and audio
- Outputs video with native stereo audio
- Up to 15 seconds at 2K resolution
- Text-, image-, audio-to-video plus editing and reference tasks
- API: $0.08/s (768P) to $0.13/s (2K) of output
AccessHailuo AI web app; MiniMax API; Hugging Face lists MiniMaxAI/MiniMax-H3 (card not opened here)
16 Jul 2026Audio/voice model
MiniMax Music 3.0
MiniMax's music generation model. It composes, arranges and sings complete songs up to five minutes long from a creative brief, optional lyrics and section tags. It reached the API in July 2026, and MiniMax followed in August with an open-weights release that it describes as production-ready.
- Full songs up to five minutes
- Vocal control: timbre, delivery and effects
- Optional lyrics and section tags as input
- Components: 8B global LLM, 0.6B local LLM, 2.4B flow module
AccessOpen weights on Hugging Face (MiniMaxAI/MiniMax-Music3) and MiniMax Audio; the paid Music API stopped taking new users on 20 Aug 2026 (pricing page lists Music-3.0 as discontinued).
1 Jun 2026Open-weights LLM
MiniMax-M3
MiniMax's current flagship: a natively multimodal mixture-of-experts model for agentic reasoning, tool use, coding and long-context work. It reads text, images and video with a context of up to 1M tokens. MiniMax published the weights on Hugging Face under its own community licence.
- ~428B total / ~23B active parameters (MoE)
- Up to 1M-token context (512K guaranteed minimum)
- Text, image and video input
- SWE-bench Verified 80.5%, BrowseComp 83.5 (lab-reported)
- Sparse attention: 9x prefill, 15x decode vs M2 at 1M
AccessMiniMax API (platform.minimax.io), MiniMax Agent; open weights on Hugging Face/GitHub under the minimax-community licence
Price$0.30 input / $1.20 output per 1M tokens (standard tier, input up to 512K)
18 Mar 2026Open-weights LLM
MiniMax-M2.7
An agent-focused update to the M2 line. MiniMax says the model took part in its own development by optimising its training harness. It targets software engineering, multi-agent 'agent teams' and professional office work, and a faster highspeed variant is sold at twice the price.
- SWE-Pro 56.22%, Terminal Bench 2 57.0% (lab-reported)
- Native multi-agent collaboration (agent teams)
- 229B parameters (Hugging Face card)
- Highspeed variant: $0.60/$2.40 per 1M tokens
AccessMiniMax API, MiniMax Agent; weights on Hugging Face (licence listed as 'other')
Price$0.30 input / $1.20 output per 1M tokens
12 Feb 2026Open-weights LLM
MiniMax-M2.5
A coding and agent model that succeeds M2.1. MiniMax says it finishes tasks 37% faster than its predecessor and costs about $1 to run for an hour at 100 tokens per second. It also targets search, tool calling and office documents, and ships with a faster variant.
- SWE-Bench Verified 80.2%, Multi-SWE-Bench 51.3% (lab)
- BrowseComp 76.3% (lab-reported)
- Programming in 10+ languages
- Office work: Word, PowerPoint, Excel tasks
AccessMiniMax API (now marked legacy); Hugging Face lists MiniMaxAI/MiniMax-M2.5
Price$0.30 input / $1.20 output per 1M tokens
23 Jan 2026Audio/voice model
MiniMax Speech 2.8
MiniMax's text-to-speech model. It adds sound tags for natural fillers such as 'um' and 'ah', clones a voice from about 10 seconds of audio, and reduces accent bleed when switching languages. It is sold through the MiniMax API, with a public demo on MiniMax's audio page.
- Sound tags for fillers and non-verbal sounds
- Voice cloning from a 10-second sample
- Background-noise and artifact reduction
- Cross-lingual output with less accent bleed
AccessMiniMax API (platform.minimax.io); demo at minimax.io/audio
23 Dec 2025Open-weights LLM
MiniMax-M2.1
A coding-focused update to M2 with stronger multi-language programming (Rust, Java, Go, C++, Kotlin, Objective-C, TypeScript, JavaScript) and native Android and iOS work. MiniMax says it gives shorter answers than M2 and uses fewer tokens. The weights are on Hugging Face.
- Multi-language coding incl. native Android/iOS
- VIBE benchmark 88.6 aggregate (lab-reported)
- More concise output than M2
- 229B parameters (Hugging Face listing)
AccessMiniMax API (now legacy); open weights on Hugging Face
Price$0.30 input / $1.20 output per 1M tokens
27 Oct 2025Open-weights LLM
MiniMax-M2
A compact mixture-of-experts model for coding and agent workflows that interleaves reasoning with tool use. MiniMax released the weights under a modified MIT licence, priced the API at $0.30/$1.20 per million tokens, and made API calls free until 7 November 2025.
- 230B total / 10B active parameters (MoE)
- Modified-MIT licence; weights on Hugging Face
- About 100 tokens/s inference (lab-stated)
- Interleaved thinking for tool use and coding
AccessMiniMax API, MiniMax Agent; open weights on Hugging Face (modified-MIT licence)
Price$0.30 input / $1.20 output per 1M tokens