What is MiniMax and how it can help you be more productive
MiniMax isn’t just the lab behind the Hailuo video generator anymore. What started in December 2021 as a Shanghai team of former SenseTime researchers has grown into one of China’s strongest open-weight challengers: a 428-billion-parameter multimodal flagship with a 1M-token context window, an agent platform priced an order of magnitude below US frontier labs, and – since its January 2026 Hong Kong IPO – a publicly listed one. In June 2026 the M3 model combined sparse attention with native image and video input, and on October 19, 2026 the M3.1 Flash halved the price again to become the default model inside MiniMax Code.
In this guide we explain what MiniMax is in 2026, how the M3, M3.1 Flash, M2, and M1 models fit together, what MiniMax Code, the mcode CLI, and Agent Teams can do, what the plans cost, and how to use it well – for coding, video, research, and everyday productivity.
What is MiniMax?
MiniMax (MiniMax Group, Chinese: XiYu Technology) is a Shanghai-based AI company founded in December 2021 by Yan Junjie and former SenseTime computer-vision researchers, named after the minimax algorithm from game theory. It is one of China’s “AI Tigers”, raised a $600 million round led by Alibaba in March 2024, and listed on the Hong Kong Stock Exchange on January 9, 2026 – raising about $619 million, with shares roughly doubling on debut. Western audiences know it best for the Hailuo video generator and the Talkie character app; developers know it for the M-series models and one of the cheapest capable coding agents on the market.
The defining trait is the combination of multimodal products, open weights, and aggressive pricing. Most flagship models ship with downloadable weights you can self-host and fine-tune, the API speaks the Anthropic protocol so existing coding tools work unchanged, and the MiniMax Code desktop app plus the open-source mcode CLI turn the models into a full agent workspace. Note the licence before shipping anything: M3-family weights use the MiniMax Community License, not MIT or Apache 2.0.
Today, depending on the surface, MiniMax can:
- Answer questions & explain concepts
- Write, debug, test & ship code (MiniMax Code)
- Reason with tunable thinking depth
- Process 1M-token codebases & documents
- Coordinate Agent Teams on long tasks
- Build persistent Mini Apps
- Generate 2K video with stereo sound
- Narrate in 40 languages & clone voices
- Compose full songs from lyrics
- Self-host open weights on your own infra
M3.1 Flash, M3, and the M-series line
MiniMax iterates fast, and retirement comes quickly: the M2.5 and M2.7 models closed to new users on October 19, 2026 and retire fully on November 18, 2026, with M Plan requests to those IDs auto-routing to M3.1 Flash. The current pairing is one flagship and one cheap workhorse.
Behind the current pair sit M2.1 (December 2025, the multilingual-coding update for Rust, Java, Go, C++, and Kotlin), M2 (October 27, 2025 – the breakout 230B/10B agent model with interleaved thinking, around 69.4% on SWE-Bench Verified in early coverage), M1 (June 2025 – the first open-weight large-scale hybrid-attention reasoning model at 456B/45.9B with 1M context), MiniMax-01 (January 15, 2025 – Text-01 and VL-01, where the Lightning Attention long-context story started), and the abab series that goes back to 2022. Beyond text, the same platform ships Hailuo / H3 for video, Speech 2.8 for voice, Music 3.0 for songs, and Image-01 for pictures.
How does MiniMax work?
MiniMax models are Mixture-of-Experts transformers: M3 activates only about 23 billion of its 428 billion parameters per token, which is how that size stays this cheap. The long-context story rests on two attention designs – Lightning Attention (linear-time attention on MiniMax-01 and M1) and MiniMax Sparse Attention on M3, where each query attends only to the most relevant key-value blocks, cutting 1M-context compute to roughly a twentieth of dense attention. Thinking is always on and tunable: M3.1 Flash runs low, medium, high, xhigh, or max effort.
A typical request flows through several stages:
- 1
Understand the request
Reads the prompt, conversation history, uploaded files, images, or video, plus connected MCP servers, skills, and persistent project memory.
- 2
Reason between tool calls
Interleaved thinking runs reasoning between tool invocations; on M3.1 Flash you tune the depth from low to max – and max costs several times more, so start low on routine edits.
- 3
Decide which tools to use
File edits, shell commands, tests, web search, browser preview, media generation, MCP servers – or splitting the work across an Agent Team with producer and verifier roles.
- 4
Generate the response or take the action
The answer, the cited report, the slide deck, the 2K video clip with sound, or a finished multi-step coding task – sometimes after working independently for minutes or hours inside MiniMax Code.
- 5
Keep the session coherent
Context compaction, checkpoints, Plan mode, and persistent memory hold project conventions across sessions, so long repository work does not reset every chat.
What can you do with MiniMax?
The agentic shift: MiniMax Code and mcode
MiniMax’s biggest shift is not longer context – it is moving from “answer my question” to “finish this workflow.” MiniMax Code began as MiniMax Agent (June 2025), became a desktop app in early 2026, was renamed in May 2026, and rebuilt on the open-source Pi Agent framework in August. Today it is two products on one subscription: a closed desktop app and an open-source terminal CLI.
An independent October 2026 harness comparison (Anchor Terminal) scored MiniMax Code 66.1 on agent readiness – ahead of Claude Code (53.6) and Gemini CLI (64.8), behind OpenCode (68.8) and Qwen Code (68.7). The strategy is the usual MiniMax play: capable enough harness, open enough code, priced so sustained agent runs stay affordable.
The MiniMax timeline: 2021–2026
Yan Junjie and former SenseTime researchers found MiniMax in Shanghai; the name comes from the minimax algorithm in game theory.
The abab model series launches (abab1 through abab3), and the Glow character-chat app reaches 5 million users in four months.
Talkie (international) and Xingye (China) relaunch the character business; Alibaba leads a $600 million round in March 2024; abab6.5 brings MoE with 200K context in April 2024.
Video-01, the first Hailuo video model, puts MiniMax on the global map with realistic motion and character animation.
MiniMax-01 (Text-01 + VL-01) introduces Lightning Attention and a 1M-token training context, extrapolating to 4M at inference – the moment MiniMax becomes a recognised open-weight lab.
M1 ships as the first open-weight large-scale hybrid-attention reasoning model (Apache 2.0), and the MiniMax Agent web product launches alongside it. Hailuo 02 follows at native 1080p.
M2 arrives as the breakout coding-and-agents model (230B/10B active, interleaved thinking, open weights) – the model that turned MiniMax into a serious Claude-Code-on-a-budget option.
M2.1 (Dec 23) strengthens multilingual coding; M2.5 (Feb 12) posts a vendor-reported ~80.2% on SWE-Bench Verified; M2.7 (Mar 18) introduces Agent Teams and self-evolution claims.
MiniMax lists on the Hong Kong Stock Exchange, raising about $619 million with shares roughly doubling on debut.
The desktop app is renamed MiniMax Code (May 29); M3 ships June 1 with sparse attention, native image and video input, computer use, and a 1M-token window.
H3 video generates 2K with native stereo sound (Jul 31); Music 3.0 weights open up (Aug 13); the mcode CLI is open-sourced (Sep 18); the Token Plan becomes the M Plan (Sep 30).
M3.1 Flash goes GA on October 19 at $0.15/$0.60 with Mini Apps landing October 22 and Flash weights public October 23; M2.5 and M2.7 begin their retirement path, fully retiring November 18.
How to use MiniMax well
The prompt pattern that works
Context, task, format – with a 1M-token window, hand MiniMax Code the whole repo. Keep effort at low until a fix fails, and verify vendor benchmark claims on your own codebase rather than trusting the marketing table.
How large is the context window?
M3 and M3.1 Flash carry a 1,000,000-token context window – roughly 750,000 words – with output up to 512K tokens. The M2 line tops out around 205,000 tokens, while M1 and MiniMax-01 were trained at 1M and extrapolate to 4M tokens at inference. Sparse and lightning attention are what keep those windows affordable to run.
Practical rule of thumb: a token is roughly 0.75 English words, and about 1.5 tokens per word once you factor in punctuation and spacing. Thinking, Agent Teams, and tool output all consume part of the window, and attention is not perfectly uniform across it – expect some “lost in the middle” degradation on very long documents.
What does MiniMax cost?
| Plan | Price | Model access | Limits |
|---|---|---|---|
| M Plan Go | $22/mo ($220/yr) | M3.1 Flash, MiniMax Code on web, desktop, CLI, and mobile; Subscription Key for Claude Code, Codex, and Cursor | Base allowance on 5-hour and weekly windows; no H3 video |
| M Plan Explore | $55/mo ($550/yr) | Everything in Go plus H3 video | 3x Go's usage |
| M Plan Build | $132/mo ($1,320/yr) | Everything in Explore plus peak-hour priority | 7.5x Go's usage |
| API · M3.1 Flash | $0.15 / $0.60 | per 1M tokens in/out, 1M context | $0.015 cached input · tunable thinking · flat at any prompt length |
| API · M3 | $0.30 / $1.20 | per 1M tokens in/out (up to 512K input) | $0.06 cached input · multimodal flagship |
| API · Speech 2.8 | $60 Turbo / $100 HD | per 1M characters | TTS, cloning, 40 languages |
| API · H3 video | ≈$0.13/sec at 2K | 2K video with native stereo, 4–15 seconds | Per-second billing; first 5 reference images free |
| Self-hosted | Free (licence) | Full weights from Hugging Face / ModelScope | MiniMax Community License conditions apply; hundreds of GB of hardware is on you |
M Plan annual billing counts 10 months for 12, and quotas run on 5-hour rolling plus weekly windows without carry-over. Explore and Build include H3 video; Go does not. Token prices are not task prices: measure cost per completed task, since a long agent run can burn far more tokens than a chat turn. The chart reflects November 2026 – check platform.minimax.io for current rates.
Frequently asked questions
Heads-up:MiniMax moves unusually fast: models retire within weeks, the agent app was renamed twice in a year, and pricing was reshaped in September 2026. The details above reflect early November 2026 (M3.1 Flash GA Oct 19; M2.5/M2.7 retire Nov 18) – check platform.minimax.io and the Hugging Face model cards for current availability, licences, and limits.