What Is MiniMax M3? MiniMax's 1M-Context Multimodal Flagship

MiniMax M3 is a natively multimodal, open-weight language model from the Chinese lab behind the M-series text models, the Hailuo video models and the Music series. The lab announced it on 1 June 2026 and it took the top slot in the company's own model list from MiniMax M2.7, bringing a one-million-token context window, native text-image-and-video input, and a sparse-attention architecture the lab calls MSA.

Fourteen articles in this blog's archive mention MiniMax M3. Until now, not one of them was about it. The model turns up as the opponent in comparison after comparison — the open-weight reference point that a newer Gemini, Qwen, Grok or GPT gets measured against — and there has been no page here that a reader could land on to find out what MiniMax M3 actually is. This is that page.

The spec sheet

MiniMax M3 is a mixture-of-experts model with roughly 428 billion total parameters and roughly 23 billion activated per token, according to the model card MiniMax publishes alongside the weights. It is natively multimodal rather than text-only with vision bolted on: the card describes training on interleaved modalities from the first step, and the model accepts text, image and video input and returns text.

• Context window: 1,048,576 tokens, with our own catalogue entry stating a guaranteed minimum of 512,000 tokens of usable context
• Maximum output: 512,000 tokens on our catalogue page; MiniMax does not publish a maximum-output figure in its own material, so treat 512K as our number, not the vendor's
• Modality: text, image and video in; text out
• Parameters: ~428B total, ~23B active per token
• Reasoning control: a thinking parameter with enabled, adaptive and disabled modes
• Recommended sampling: temperature 1.0, top_p 0.95
• Architecture: MiniMax Sparse Attention (MSA), the subject of arXiv paper 2606.13392
• Endpoint shape: OpenAI-compatible chat completions

The headline architectural claim is about attention cost at long context. MiniMax says MSA delivers more than 9x faster prefilling and more than 15x faster decoding than M2 at a one-million-token window, cutting per-token compute to roughly one-twentieth of the previous generation. Those are vendor figures and we have not seen them independently reproduced.

Where MiniMax M3 sits in MiniMax's own line

It is the top of the language line, but it is worth being precise about what that line contains, because MiniMax keeps a longer tail of still-listed models than most labs. The vendor's own model documentation splits them in two.

• Current models: MiniMax M3, MiniMax M2.7, MiniMax M2.7-highspeed
• Legacy models, still listed: MiniMax M2.5, MiniMax M2.5-highspeed, MiniMax M2.1, MiniMax M2.1-highspeed, MiniMax M2
• MiniMax M2, the oldest of them, carries a 200,000-token context and a 128,000-token maximum output including chain-of-thought
• MiniMax does not publish context or output limits for M2.7 or M2.1 in the same document

The generational gap is real but not enormous on the one index both models appear on. Artificial Analysis scores MiniMax M3 at 29 on the Intelligence Index revision v4.3.2, and MiniMax M2.7 at 23 on that same revision — 29 puts M3 at rank 16 of 114 models in that revision's table, and 23 puts M2.7 at rank 36. M2.7 was released in March 2026. The jump is a step, not a generation, and it is worth reading the two numbers off the same revision: this index was recalibrated on 7 September 2026, and scores from before that date are not comparable with scores after it.

What it costs

MiniMax prices M3 by request size on its pay-as-you-go rate card, and the published rates carry a permanent 50% discount against a higher list price.

• Up to 512K tokens of input: $0.30 per million input tokens, $1.20 per million output tokens, $0.06 per million cached reads — against a list of $0.60 / $2.40 / $0.12
• Above 512K tokens of input: $0.60 per million input, $2.40 per million output, $0.12 per million cached reads — against a list of $1.20 / $4.80 / $0.24
• Priority service tier: 1.5x the standard rate, so $0.45 / $1.80 / $0.09 on the small-request tier, selected by setting service_tier to "priority"
• Not published: MiniMax lists no cache-write price for M3, only a cache-read price

OrcaRouter carries MiniMax M3 at the same numbers — $0.30 in, $1.20 out — with no markup on top of the provider rate. It is a featured model in our catalogue, and the traffic is not trivial: 153.7 million tokens routed in the last seven days at the time of writing, with a p50 time-to-first-token of 4.26 seconds and a p95 of 10.00 seconds.

What MiniMax M3 is measurably good at

Start with the numbers MiniMax reports for itself, all of which are vendor-reported and none of which we have seen reproduced by a third party.

• SWE-Bench Pro: 59.0%
• Terminal-Bench 2.1: 66.0%
• SWE-fficiency: 34.8%
• MCP Atlas: 74.2%
• BrowseComp: 83.5, the figure the vendor leads with on agentic search
• Video-MME: 84.6 at 512 frames, where the external API caps frames at 640
• KernelBench Hard: 28.8%
• OSWorld-Verified: 68.70% at 100 maximum steps, rising to 70.06% at 200

The independent picture agrees on the direction. Artificial Analysis puts MiniMax M3 at an Intelligence Index of 29 on revision v4.3.2, ranking 16th of 114 models, at $0.51 to run the index per model — rank 21 of 114 on that cost measure. Output speed is 106.4 tokens per second against a 67.1 median, and time to first token is 1.25 seconds against a 2.33 second median. Both of those are better than typical for the tier, and the price is well below it.

Third-party republishing of the same index family fills in the sub-scores: SWE-bench Verified at 80.5%, tau-squared-bench at 88.9%, AA-GPQA Diamond at 92.9%, AA-LCR at 83.0%, AA-IFBench at 82.9%, OmniDocBench 1.5 at 91.6%, and USAMO 2026 at 85.7%. Document understanding and instruction following look like genuine strengths, and the long-context reasoning score matches the architectural story.

What it is measurably bad at

The honest weaknesses cluster in two places, and both are visible in published numbers.

• The agentic gap is the big one: on the same third-party republishing, MiniMax M3 scores 58.6 on the AA Coding Index but only 30.8 on the AA Agentic Index. It writes code far better than it drives a long multi-step task autonomously.
• Knowledge and hallucination: an AA-Omniscience Index of 1.4, accuracy of 16.7%, and a hallucination rate of 18.4%. This is a model that will answer confidently about things it does not know.
• OSWorld 2.0: 4.6%
• CritPt: 3.7%, the weakest published score we found for M3 anywhere
• ResearchClawBench: 19.8%
• Verbosity: 120 million output tokens to complete the Intelligence Index, rank 11 of 114 — it is a talkative model, and on reasoning-heavy prompts that shows up on the bill
• The composite floor: one third-party aggregator puts it at 55.28 out of 100, rank 60 of 505, though on partial coverage of 50 benchmarks out of 481, so that composite is a floor rather than a verdict

Several things are simply not measured, and we would rather say so than fill the gap. Nobody outside MiniMax has reproduced the vendor benchmark table above. Artificial Analysis does not publish a Coding Index or an Agentic Index figure for M3 in its own right — the 58.6 and 30.8 pair comes from a third party republishing the same index family. Artificial Analysis names AA-Omniscience as a component of its index but publishes no numeric Omniscience score for M3. MiniMax publishes no maximum-output figure, so the 512K on our own model page is ours. And there is no cache-write price on the vendor rate card at all.

Who should pick MiniMax M3, and who should pick something else

Pick MiniMax M3 when the job is long-context and multimodal at once. One million tokens with video input, open weights you can download and run yourself, and an input price of $0.30 per million tokens is a combination nothing else in the open-weight field matches cleanly. Long-document analysis, video understanding, repository-scale code reading and document-extraction pipelines are where the measured strengths sit — 91.6% on OmniDocBench and 83.0% on AA-LCR are the numbers that argue for it.

Pick something else in three cases. If you do not need the context window or the vision input, MiniMax M2.7 costs the same $0.30/$1.20, scores 23 against M3's 29 on the current index revision, and is a smaller thing to operate — M3's advantage is not free, it is just priced at the same sticker. If your workload is autonomous agents rather than coding, the agentic gap is the reason to look elsewhere, and MiniMax's own comparison table points at the same place: on PostTrainBench the vendor reports 0.37 for M3 against 0.42 for Opus 4.7 and 0.39 for GPT-5.5, its only published benchmark where M3 loses to both. And if you need a licence with no strings, read the next paragraph before you commit.

The licence is the catch that a lot of coverage skips. MiniMax M3 ships under the MiniMax Community License, not Apache or MIT. Non-commercial use is free. Commercial use is permitted, but you must prominently display "Built with MiniMax M3". Below $20 million in annual revenue you file a one-time notice; above it you need prior written authorization from MiniMax, requested by email with the subject "M3 licensing - authorization request". The licence also carries an appendix of prohibited uses. If you are shipping a product, that is a legal review, not a formality.

The practical summary

MiniMax M3 is the current flagship of MiniMax's language line: a ~428B-parameter mixture-of-experts model with ~23B active per token, a one-million-token context, native video and image input, and sparse attention that the vendor credits with a 20x reduction in per-token compute at full context. It was announced on 1 June 2026, so this is a settled model, not a new one — the interesting question in September 2026 is not whether it is good, it is which half of your workload it is good for.

The measured answer is that it is strong at code and documents, ordinary at knowledge, and weak at autonomous agentic work. It is cheap for what it does, and it is the rare open-weight model where the long-context claim survives an independent look. The licence is the part most teams will need to negotiate with themselves.

MiniMax M3 runs on OrcaRouter at the provider rate with zero markup, through the same OpenAI-compatible endpoint as everything else here: one API key reaches 200+ models, with automatic failover if an endpoint drops and a routing DSL when you want to pin or blend. If you want to compare it against the rest of the catalogue before you commit, the model page has the live rate card, time-to-first-token percentiles and traffic.