What is MiniMax M3?
MiniMax M3 is a native multimodal mixture-of-experts model with about 428 billion parameters in total and about 23 billion activated per token, released as open weights under the MiniMax Community License. Its defining component is MiniMax Sparse Attention, an attention design the weight card credits with a nine-fold prefill and fifteen-fold decode speedup against the previous generation at a million-token context, and with per-token compute cut to one twentieth. The context window is 1,000,000 tokens, and the product page adds that the API guarantees a minimum of 512,000 of them, which is a statement about the floor of the window rather than a service-level commitment. Input is native text, image and video and there is no audio input, and the vendor describes output at around 100 tokens per second. The weights are published as an MXFP8 quantised build of roughly 444 gigabytes alongside a BF16 build of roughly 854 gigabytes, and the deployment guidance is published with them as an experimental reference rather than a supported production configuration. Three relationships are worth fixing before the cards below. First, the number that describes the model and the number that describes one reply differ by more than an order of magnitude, and only the smaller one is visible in this workspace. Second, open weights and an open licence are two separate facts: the download is free, and commercial use carries an attribution requirement and a revenue threshold. Third, MiniMax ships a Flash-Preview tier beside this model and a previous generation behind it, and nothing about either of them is reported here.
What MiniMax M3 is worth using for
A million-token window with a guaranteed floor of 512,000, and no counter anywhere on this page
The window is the headline of this model and it is also where this page diverges from the documentation in the most easily missed way. MiniMax documents 1,000,000 tokens, and the product page adds that the API guarantees a minimum of 512,000, which matters because it says the window is not merely advertised at a million but has a published floor beneath it. The subscription behind the API bills differently on either side of that floor, which the pricing section below handles. What this workspace shows you is nothing at all: the catalogue it reads carries no context field for this model, so no usage counter or window figure appears anywhere in the chat box, and every number you read here is the documented ceiling rather than a live measurement of your session. That is a property of how this site reads its catalogue rather than a limitation of the model, and it applies to every model on this site the same way. The practical consequence is that long documents are supported and unmeasurable at the same time.
Here is a long technical report and a question about one table inside it. Answer from the table, quote the paragraph around it, and tell me whether your answer depended on any part of the document that did not fit.
Text, image and video in the documentation, and a file picker here that takes images and PDFs
This is the first page in the text-model set whose subject both documents image input and actually accepts it here, so the two sides of this card agree more than they do on the neighbouring pages. MiniMax documents text, image and video input with no audio support, and the API reference puts numbers on each: an image may be up to 10 megabytes in JPEG, PNG, GIF or WEBP, a video may be up to 50 megabytes through a URL or a base64 body or up to 512 megabytes through the file interface, in MP4 or MOV, and the whole request body may not exceed 64 megabytes. Video frames are sampled at a rate that defaults to 1 and accepts values from 0.2 to 5, and an image carries a detail setting that takes low, default or high. The picker in the chat box above accepts JPEG, PNG and WEBP images and PDF documents up to 24 megabytes, and it does not accept video. So the image half of the documented capability is reachable here and the video half is not, which is the reverse of the caveat the two MiMo pages above have to carry.
I am uploading a screenshot of a failing dashboard and the specification document it is supposed to implement. Compare the two, list every place they disagree, and rank the disagreements by how much they would cost to fix.
Thinking that can be switched off, a split between reasoning and answer, and a rule for multi-turn tool calls
Thinking on this model is a request-level choice rather than a model choice, and the vendor documents it through a field that accepts two values. Setting it to the value that disables thinking skips reasoning and answers directly, which is the configuration this card is about; the other documented value leaves the amount of reasoning to the model. The model also gives you a way to keep reasoning out of the answer text, because a boolean on the request moves reasoning into its own response field, and leaving that setting alone returns reasoning wrapped in tags inside the normal content. One rule matters for anything agentic: the model reasons between tool calls rather than once at the start, and a multi-turn conversation has to send the complete assistant message back, tool calls and reasoning included, or the model loses the thread it reasoned along. None of this is reachable from the box above, because the catalogue registers no reasoning capability for this model and the control is therefore disabled rather than absent.
Answer this question twice in the same conversation: once with the reasoning included in the reply, once with the reasoning separated out. Then tell me which of the two you would keep in a shared transcript and why.
Open weights under a community licence, and a 444-gigabyte download before you can run them
The weights are downloadable and the licence is the part to read before anything commercial. MiniMax publishes a quantised build of about 444 gigabytes and a BF16 build of about 854 gigabytes, and the deployment guidance gives an eight-GPU single-node reference configuration at a tensor parallelism of eight. Two things attach to that guidance and both are easy to skip. It is published as an experimental reference rather than a supported production configuration, and the vendor says explicitly that it should not be read as a service-level commitment. And an activated-parameter count is not a memory figure: the same page warns that a model described by its activated parameters still needs its full weights resident, or sharded across devices. The licence is the MiniMax Community License, which is not MIT and not Apache. Non-commercial use is free, commercial use requires the attribution notice Built with MiniMax M3 to be displayed prominently, and an organisation above a published annual-revenue threshold has to obtain written authorisation first.
I am planning to ship a product that calls this model through a hosted provider and also fine-tunes the weights locally. Walk through what the licence requires in each case, mark which requirements are obligations rather than suggestions, and list anything you would ask a lawyer rather than me.
Official MiniMax M3 specifications
- Identifier
- minimax-m3 — the identifier this catalogue uses; the official model name is MiniMax-M3
- Parameters
- About 428 billion total, about 23 billion activated per token — a native multimodal mixture of experts
- Architecture
- MiniMax Sparse Attention: nine-fold prefill and fifteen-fold decode speedups against the previous generation at a 1M context, with per-token compute cut to one twentieth
- Context window
- 1,000,000 tokens, with a guaranteed minimum of 512,000
- Maximum output
- 131,072 tokens recommended and 524,288 maximum, through max_completion_tokens
- Input modality
- Text, image and video in; text out. No audio input
- Thinking
- thinking.type accepts disabled or adaptive; disabled skips thinking and answers directly
- Sampling parameters
- temperature defaults to 1 within a range of 0 to 2, and top_p defaults to 0.95
- Long-context threshold
- Input above 512,000 tokens moves to the long-context price tier, cache-hit tokens included
- Licence
- MiniMax Community License — commercial use must display the notice Built with MiniMax M3 prominently
- Open weights
- MiniMaxAI/MiniMax-M3-MXFP8 at about 444 GB; the BF16 build is about 854 GB
- API protocols
- OpenAI Chat Completions, Anthropic Messages and OpenAI Responses officially; this catalogue registers an OpenAI-compatible endpoint only
MiniMax M3 and Claude Opus 4.7 on the vendor’s own page
Every value below is read from one page — the MiniMax M3 product page — and only two rows on that page carry a number on both sides, which is why this table is short. The page makes its coding and agentic case through charts rendered as images, with no numeric labels to copy, so those results cannot be reproduced here even though they are the vendor’s headline; the interactive tabs beside them, labelled benchmark and architecture and long-horizon stability, are the same. The two rows that do carry readable values are an autonomous-browsing benchmark, where the page states the M3 score and the rival score in a single sentence, and an autonomous-training benchmark, where it states the M3 score, its own rank and two rival scores. That second row goes against the model and it is kept rather than dropped: on the training benchmark the page places M3 third behind two named rivals, and the row above it is the only one the model wins. One caveat applies to both. These are the vendor’s own measurements of its own model against rivals it chose, not an independent evaluation, and neither row predicts the quality of your conversation.
| Official benchmark | MiniMax M3 | Claude Opus 4.7 |
|---|---|---|
| BrowseComp · Agentic | 83.5 | 79.3 |
| PostTrainBench · Agentic | 37.1 | 42.4 |
How to use MiniMax M3 online
Pick the half of this model you are actually going to use
The documented capability has two halves and this workspace carries one of them fully. If your workload is long text, use the window: a million tokens are documented and a floor of 512,000 is guaranteed, and the price tier changes above that floor, so the length of your prompts is a billing decision as well as a technical one. If your workload is visual, images are reachable here and video is not, because the picker accepts JPEG, PNG and WEBP images and PDF documents up to 24 megabytes and refuses video outright. Deciding which half you need before you build is more useful than reading the specification in full, because the specification promises both and this interface delivers one.
Treat the two output numbers as different objects
There are two ceilings on a reply and they come from different places. The vendor documents a recommended reply length of 131,072 tokens and a maximum of 524,288, and both of those are properties of the model’s API. The chat box above requests at most 4,096 output tokens per reply, because the composer applies its own ceiling to whatever value the catalogue publishes. The catalogue value is not wrong and the composer value is not a bug; they answer different questions. What matters when you plan is which one binds, and here it is the smaller one, so a task that depends on a single very long answer is better broken into parts than described in full.
Read the licence before you read the price
This is an open-weight model, which makes it easy to assume a permissive licence, and that assumption is wrong here. The weights ship under a community licence: non-commercial use is free, commercial use must display a specific attribution notice prominently, and an organisation above a published revenue threshold must obtain written authorisation before it ships. The download and the hosted API are not restricted in the same way, so the two routes have to be checked separately. If you are building anything commercial on top of this model, the licence text is the first document to read and the price table is the second.
Read the price as three meters on two tiers
The published price is not one number. Input, output and cache reads are each metered separately, and input crosses into a second and more expensive tier once a request passes the long-context threshold, with cache-hit tokens counted towards it. On top of that the vendor currently publishes a discounted standard tier beside a struck-through original and marks the reduction as permanent rather than promotional, which is unusual enough to be worth checking again later, and a priority service tier exists at a published multiple of the standard rates. Nothing above is this site’s checkout price: this site bills in points against its own catalogue reference price, and the two are not convertible.
What MiniMax M3 costs, and which official number belongs to which meter
One official table covers this model and it has three columns and two tiers. For input up to the long-context threshold the standard rates are thirty cents per million input tokens, one dollar twenty per million output tokens and six cents per million cache reads. Above the threshold the same three columns rise to sixty cents, two dollars forty and twelve cents. The table prints those standard rates beside struck-through original figures and marks the difference as a permanent discount rather than a limited promotion, which is why this page repeats the discount wording instead of the original numbers. Two structural details come from the same artefact. The price is metered on three separate columns, so a workload that reads a large cached prefix and writes a short answer is not priced by any single rate. And the threshold is applied to the request rather than to the account, with cache-hit tokens counted towards it. A priority service tier is published at one and a half times the standard rates and is selected through a request field. One column a reader might expect is absent: the official table gives no cache-write price for this model. Nothing above is this site’s checkout price. This site bills in points against a catalogue reference price of 136 points per million input tokens, 536 points per million output tokens and 100 points per million cache reads, and points are not a conversion of dollars.
View points and plans ↗MiniMax M3 FAQ
How much context does MiniMax M3 have?
A million tokens, and the product page adds a guarantee that a minimum of 512,000 tokens of that window is available, which is a statement about the floor rather than a service-level commitment. This interface displays no context figure at all, because the catalogue it reads carries no context field for this model, so every number you see quoted for this page is the documented ceiling rather than a measurement of a live session. The window also interacts with price rather than only with capability: input above the 512,000-token threshold falls into a different tier, cache-hit tokens included, which the pricing section above sets out. If your prompts are long enough to reach that tier, the threshold is worth knowing before you design the request rather than after.
Does it accept images and video?
The documentation says text, image and video, and it does not list audio. The API reference puts limits on each: an image may be up to 10 megabytes in JPEG, PNG, GIF or WEBP, a video may be up to 50 megabytes through a URL or a base64 body or up to 512 megabytes through the file interface in MP4 or MOV, and the whole request body may not exceed 64 megabytes. Image requests carry a detail setting, and video frames are sampled at a rate that defaults to 1 and accepts values from 0.2 to 5. This workspace is narrower in one direction and wider than its neighbours in another: the file picker accepts JPEG, PNG and WEBP images and PDF documents up to 24 megabytes, and it does not accept video at all.
Can thinking be switched off?
Yes, through a request field whose disabling value skips thinking and answers directly, while the other documented value leaves the amount of reasoning to the model. The model can also be told to keep reasoning out of the answer text, because a separate boolean moves it into its own response field, while leaving that setting alone returns reasoning wrapped in tags inside the normal content. Multi-turn tool use has a rule of its own: the model reasons between tool calls, so a follow-up request has to carry the complete assistant message back, tool calls and reasoning included. The control in this workspace is disabled rather than hidden, because the catalogue registers no reasoning capability for this model.
How long can one reply be?
Two answers apply and they differ. The vendor documents a recommended reply length of 131,072 tokens and a maximum of 524,288, both as properties of the model’s API. The chat box above requests at most 4,096 output tokens for each reply, because the composer applies its own ceiling to whatever the catalogue publishes. Neither number is an error; they describe different layers. For planning purposes the smaller one binds here, which means a task that needs a single very long answer is better broken into several requests than described in one.
Is MiniMax M3 open source?
The weights are open and the licence is not a permissive one. MiniMax publishes a quantised build of about 444 gigabytes and a BF16 build of about 854 gigabytes, with the technical report and the code repositories linked from the weight card, and it names three inference frameworks it recommends. The licence is the MiniMax Community License rather than MIT or Apache. Non-commercial use is free, commercial use must display a specific attribution notice prominently, and an organisation above a published annual-revenue threshold must obtain written authorisation before shipping. The official deployment guidance is published as an experimental reference and says so in the same breath, and it warns that a model described by its activated-parameter count still needs its full weights resident.
What does MiniMax M3 cost?
There are three meters and two tiers. For input up to the long-context threshold the published standard rates are thirty cents per million input tokens, one dollar twenty per million output tokens and six cents per million cache reads; above that threshold they rise to sixty cents, two dollars forty and twelve cents. The table prints those rates beside struck-through original figures and marks the reduction as permanent rather than promotional. A priority service tier is published at one and a half times the standard rates and is selected through a request field, and the official table lists no cache-write price for this model. This site does not bill in dollars: it bills in points against a catalogue reference price of 136 points per million input tokens and 536 points per million output tokens.
Does it support web search?
MiniMax supports a search tool for this model, but not through the interface this site uses. The tool is documented as a beta server-side feature billed per request rather than per token, and the vendor states that it is reachable only through two specific APIs, neither of which the catalogue behind this page registers: the endpoints declared here are OpenAI-compatible chat completions only. That is why the search control in the chat box above is present and disabled rather than simply missing. The distinction matters because a model supporting search in its own console does not imply that every caller of that model can switch search on, and the gap here sits on the caller’s side rather than the model’s.
How does it compare with Claude Opus 4.7?
On the vendor’s own product page, two rows carry a value on both sides. On an autonomous-browsing benchmark the page reports a score of 83.5 against 79.3, and on an autonomous-training benchmark it reports 37.1 against 42.4 while also placing a second rival above it at 39.3, which is why that row is a loss in this table. The page’s wider coding and agentic claims are presented as charts without numeric labels, so they are not reproduced here. Both rows are the vendor’s measurements of its own model against rivals it selected, not an independent evaluation, and neither one predicts how the model will behave on your task.
Sources and model information
Information checked
Start a MiniMax M3 conversation
The chat box above runs this model, and this page is written to be checked rather than trusted. The specifications and the benchmark rows come from the vendor’s product page, the parameter-level rules and the multimodal limits come from the API reference, the prices come from the official pay-as-you-go table, the deployment baseline and its experimental status come from the self-hosting guide, the licence terms come from the licence text shipped with the weights, and the limits of this workspace come from the code that runs it. Where the two sides differ, the difference is written out above rather than smoothed over, and where an official artefact stops short of a figure, the page says that too. The three facts here least likely to be repeated elsewhere are that this workspace can send an image but not a video while the model documents both, that the licence carried by the open weights requires an attribution notice for commercial use, and that the only benchmark rows the vendor prints with a number on both sides are two agentic evaluations rather than the coding results the page leads with.
Start a chat with MiniMax M3 ↑