What is MiniMax M3?
MiniMax M3 is a mixture-of-experts model with about 428B parameters, 23B of them active for each token. MiniMax trained it on text, images and video together from the start, and its sparse attention design lets it work across up to 1M tokens, with 512K guaranteed on the API. It can think before answering, skip thinking for speed, or decide for itself, and MiniMax says it can operate a desktop computer.
It is MiniMax's current flagship on the API. A newer MiniMax-M3.1-Flash-Preview is offered only on the Token Plan and in MiniMax Code, and the M3 weights are on Hugging Face under a community licence.
What changed from MiniMax M2.7?
M3 adds a much longer context, image and video input and a looser licence, at twice M2.7's normal API price.
- Context
- Up to 1M tokens, with 512K guaranteed, against 204,800 for M2.7.
- Input
- Text, images and video, where M2.7 reads text only.
- Licence
- Business use of the weights needs a credit and a notice, or written approval above $20 million a year; M2.7 needs written approval for any business use.
- Price
- $0.60 / $2.40 per million input / output tokens against $0.30 / $1.20 for M2.7; one Token Plan quota covers both.
Who is MiniMax M3 best for?
MiniMax M3 is best for developers running long coding and agent jobs that need a 1M-token context or image and video input, on the API, a flat monthly Token Plan or their own GPUs.
Long coding and agent sessions
The API takes up to 1M tokens of context, with at least 512K guaranteed, so a large codebase and its logs fit in one window.
Flat-rate use in coding tools
A Token Plan key works in Claude Code, Cline, Cursor and other OpenAI-compatible tools for a fixed monthly price.
Work that mixes code and visuals
M3 reads images and video natively, so it can work from screenshots, charts and diagrams.
Worker agents under a stronger planner
Several developers get their best results when a stronger model writes the plan and M3 carries out the steps at a lower token cost.
What's included with MiniMax M3?
A 1M-token context, image and video input, three thinking modes, computer use, Anthropic- and OpenAI-compatible endpoints, and downloadable weights.
- 1M-token context
- MiniMax Sparse Attention cuts the compute per token at 1M tokens to a twentieth of the previous generation's, and the API guarantees at least 512K.
- Image and video input
- Trained on text, images and video together, so it reads screenshots, charts and video alongside code.
- Three thinking modes
- Thinking can be on, off or adaptive, and on the API thinking on and off cost the same.
- Computer use
- MiniMax says M3 can operate a desktop, and MiniMax Code uses it to work across apps and files; it scored 70.06% on OSWorld-Verified in MiniMax's test.
- Two API formats
- Anthropic- and OpenAI-compatible endpoints with automatic prompt caching, so Claude Code, Cursor, Cline and similar tools can call it directly.
- Downloadable weights
- On Hugging Face, with serving guides for SGLang, vLLM, Transformers and KTransformers.
How much does MiniMax M3 cost?
MiniMax M3's weights are free to download for non-commercial use; its API costs $0.60 per million input tokens and $2.40 per million output tokens, and Token Plan subscriptions for coding tools start at $22 a month.
- MiniMax Community License on Hugging Face
- Free for non-commercial use
- Business use needs a credit and a notice; approval above $20 million a year
- Prompts over 512K tokens: $1.20 / $4.80
- Cache reads: $0.12 per million tokens
- Priority tier costs 1.5x the standard rate
- About 1.7B tokens of M3 use a month
- 3 to 4 agents at once
- Text, image and speech share one quota
- About 5.1B tokens of M3 use a month
- 4 to 5 agents at once
- 3 Hailuo 2.3 video clips a day
Token Plan Ultra costs $132 a month for about 12.5B tokens of M3 use, 6 to 7 agents at once and 5 Hailuo 2.3 video clips a day. Billed yearly, Plus is $220, Max $550 and Ultra $1,320 a year, which MiniMax counts as two months free. Token Plan quotas reset on 5-hour rolling and weekly windows and also cover MiniMax Code. Prepaid credits cost $1 per 1,000 credits in $5, $25 or $100 packs and last 365 days. Right now the page shows the MiniMax-M3 API at $0.30 / $1.20 per million input / output tokens instead of $0.60 / $2.40, with no end date. Full pricing on platform.minimax.io. We checked these prices on .
How do readers rate MiniMax M3?
No one has rated MiniMax M3 yet. Be the first.
Sign in to rate. One rating per account.
MiniMax M3 pros and cons
MiniMax M3 pairs a 1M context and video input with a low price and strong scores on MiniMax's tests, but users report uneven instruction following and its licence has strings for business use.
Pros
- Reads text, images and video across a 1M-token context, which few open-weight coding models do: GLM-5.2 is text only and MiniMax M2.7 stops at 204,800 tokens.
- Cheap at volume: $22 a month buys about 1.7B tokens of M3 use on the Token Plan, and several developers call it some of the best value they have found.
- Strong numbers on MiniMax's own agent tests: 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1 and 83.5 on BrowseComp, where Opus 4.7 scored 79.3.
- Many users say it plans and reasons well and is a clear step up from MiniMax M2.7, and several use it as a worker that carries out another model's plan.
Cons
- Users are split on instruction following: several say it ignores or forgets instructions, or gets confused and hangs in long agent runs in Hermes, Cline and OpenCode.
- Some developers rank its code below GLM or DeepSeek's Flash model, and say it is slow and uses many tokens on large changes.
- The weights are not free for business: commercial use needs a "Built with MiniMax M3" credit and a notice to MiniMax, and products earning over $20 million a year need its written approval.
- Prompts over 512K tokens cost twice as much, at $1.20 / $4.80 per million input / output tokens, and only 512K of the 1M context is guaranteed.
What are the best MiniMax M3 alternatives?
Three tools on Recaply do MiniMax M3's job. DeepSeek V4.1 Flash and GLM-5.2 are open source, so you can run them on your own server.
DeepSeek V4.1 FlashOpen SourceChoose it when price comes first: $0.30 / $1.20 per million input / output tokens at peak and half that off-peak, with a 1M context, image input and MIT-licensed weights.
GLM-5.2Open SourceChoose it when you want MIT-licensed weights and steadier from-scratch builds and only need text; it has a 1M context and costs $1.4 / $4.4 per million input / output tokens.
MiniMax M2.7FreeChoose it when text is enough and you want half M3's normal API price, at $0.30 / $1.20 per million input / output tokens, with a 204,800-token context.
Questions about MiniMax M3
What is MiniMax M3.1-Flash-Preview?
It is a newer MiniMax model with the same 1M context and five effort levels for thinking depth. For now MiniMax offers it only through the Token Plan and MiniMax Code, so on the pay-as-you-go API MiniMax-M3 is still the model to call.
Who is MiniMax M3 for?
MiniMax M3 is best for developers running long coding and agent jobs that need a 1M-token context or image and video input, on the API, a flat monthly Token Plan or their own GPUs.
Is MiniMax M3 free?
MiniMax M3's weights are free to download from Hugging Face and free to use for non-commercial work. MiniMax's API costs $0.60 per million input tokens and $2.40 per million output tokens, and there is no free Token Plan; plans start at $22 a month.
Does MiniMax M3 cost more for long prompts?
Yes. On the MiniMax M3 API, prompts over 512K input tokens cost $1.20 per million input tokens and $4.80 per million output tokens, twice the rate for shorter prompts. Cache reads rise from $0.12 to $0.24 per million tokens.
Is there an open source alternative to MiniMax M3?
Yes. DeepSeek V4.1 Flash and GLM-5.2 are open source tools that do the same job, so you can run your own.
What are the best MiniMax M3 alternatives?
The closest alternatives on Recaply are DeepSeek V4.1 Flash, GLM-5.2 and MiniMax M2.7.
Where can I try MiniMax M3?
On its official website, minimax.io.
Can I use MiniMax M3 for commercial work?
Yes, with conditions. On MiniMax's API you simply pay per token. With the open weights, MiniMax's community licence asks you to show "Built with MiniMax M3" and send MiniMax a one-time notice; products earning over $20 million a year need written approval first.
Can I use MiniMax M3 in Claude Code or Cursor?
Yes. A MiniMax Token Plan gives you a subscription key for Claude Code, Cursor, Cline, Roo Code, Kilo Code, OpenCode and other OpenAI-compatible tools, and the same quota covers MiniMax's own MiniMax Code agent. Plans start at $22 a month.
Our verdict
MiniMax M3 is good value for long, multimodal agent work, as long as a stronger model or a person checks its output.
Few open-weight models match its 1M context with image and video input, it scores well on MiniMax's own agent tests, and the Token Plan makes heavy use cheap; but users disagree on how well it follows instructions.
- Skip it for
- Teams that need a dependable lead agent for from-scratch builds, and businesses earning over $20 million a year that do not want to ask MiniMax for approval to self-host.
- Check first
- Run your own agent tasks on it before you commit, and see whether your prompts pass 512K tokens, where the price doubles.
MiniMax M2.5
Kimi K3
Muse Spark 1.1
Qwen3.8-Flash
Qwen3.6-Plus
Kimi K2.5