What it is
MiniMax-M3 is the flagship open-weight model from Shanghai-based MiniMax, released on 1 June 2026, with weights on Hugging Face from 12 June. It is a mixture-of-experts model with about 428 billion parameters, 23 billion active per token, trained on text, images and video from the first step. It reads all three, writes text and has a 1M-token context. A smaller MiniMax-M3.1-Flash now sits beside it on MiniMax's API.
Its MiniMax Sparse Attention design is built for million-token inputs; against M2 at 1M tokens, MiniMax claims 9x faster prompt processing, 15x faster generation and a twentieth of the compute per token. Thinking can be enabled, adaptive (the model decides) or disabled.
MiniMax's API charges $0.30 per million input tokens, $1.20 per million output and $0.06 for cached input on prompts up to 512K tokens, after a 50% discount it labels permanent, and double that above 512K; a priority tier costs 1.5 times as much. The weights carry a MiniMax community licence that requires a "Built with MiniMax M3" credit for commercial use and written permission above $20 million in annual revenue.
Our review
MiniMax-M3 is a capable but mid-table open model whose selling point is cheap long context, and whose licence is the reason to slow down. Artificial Analysis gives it 29 on its Intelligence Index, 23rd of 121 large open-weights models, against a class median of 19. MiniMax's own card reports 80.5 on SWE-bench Verified, 59 on SWE-bench Pro and 85.4 on Video-MME-v2; its chart against named rivals is published only as an image, so those comparisons cannot be read.
Price is where it competes. Artificial Analysis puts the blended rate at $0.22 per million tokens with caching, and measured 86.4 output tokens a second on MiniMax's API with 1.72 seconds to first token, close to the median for its class. Input past 512K tokens costs double, so a single full 1M-token prompt is about $0.60 in input alone: cheap by frontier standards, though not something to resend on every turn without the cache.
The licence is the most awkward of the major open models. Its grant is worded for "non-commercial purposes", then sets conditions for commercial use: a visible "Built with MiniMax M3" credit, a notice email to MiniMax from every commercial user, prior written permission for products earning over $20 million a year, and a ban on any military purpose. It names no governing law. GLM-5.3 and Mistral's Apache-licensed models carry none of this.
Self-hosting is a serious undertaking. At 427 billion parameters the BF16 weights come to roughly 850 GB, more than an eight-GPU H100 node holds, so most teams will reach for one of the 58 community quantisations or AMD's MXFP4 recipe, or rent it. MiniMax's API retention and training terms were not published on the pages we could read, which leaves the open weights as the only route with clear data control.
Where it earns its place
A legal-tech engineer working on M&A data rooms
Ask questions across a whole contract bundle in one prompt
Send the bundle as one long request instead of building retrieval. A 750K-token prompt costs about $0.45 in input at the above-512K rate, and cached re-reads drop to $0.12 per million, so follow-up questions on the same bundle are far cheaper than the first.
A media-monitoring start-up summarising broadcast clips
Turn hours of news video into searchable summaries
M3 reads video natively, scoring 85.4 on Video-MME-v2 by MiniMax's own report. Assuming 200K input and 2K output tokens per clip, 10,000 summaries a month is about $624 on the standard tier.
A product team at a company with $50 million in annual revenue
Ship a feature on self-hosted open weights
Above the licence's $20 million threshold they need MiniMax's written permission before use, plus the "Built with MiniMax M3" credit in the product. That approval, not the GPUs, usually sets the timeline; an Apache-licensed model avoids it.
Before you commit
- Setup
- API: create a key on platform.minimax.io and set thinking to enabled, adaptive or disabled per request. Weights: download from Hugging Face and serve with SGLang or vLLM; MiniMax recommends temperature 1.0 and top_p 0.95.
- Integrations
- MiniMax API (also on OpenRouter at $0.30/$1.20), MiniMax Agent, and open-source serving through SGLang, vLLM, Transformers, KTransformers, Unsloth and AMD's ATOM for ROCm.
- Data and privacy
- Licence: commercial use needs a "Built with MiniMax M3" credit and a notice to MiniMax; products over $20 million in annual revenue need prior written permission; no military use. MiniMax's API retention and training terms were not published on the pages we could read.
- Cost at scale
- 1M support replies a month at 2,000 input and 300 output tokens each is about $960 on the standard tier, less with cached system prompts. Self-hosting at BF16 needs roughly 850 GB of GPU memory before KV cache: an eight-GPU H200 or B200 node, or more than one H100 node.
- Limits
- 1M-token context, though OpenRouter's top provider serves 524K, and prices double past 512K. Maximum output is not stated on the card. On Long-Horizon Terminal-Bench MiniMax reports a 38.5 mean reward, solving 3 of 46 tasks at the strict threshold.
What's good
- API at $0.30/$1.20 with $0.06 cached input
- 1M-token context with sparse attention
- Text, image and video input in one model
- Runs on vLLM, SGLang, KTransformers and AMD ROCm
What's not
- Licence needs attribution, a notice and permission above $20M revenue
- Mid-table on Artificial Analysis: 29, 23rd of 121 large open models
- BF16 weights around 850 GB
- API retention and training terms not found
Best for
Cost-sensitive teams processing long documents or video who want an open model and can accept its licence terms.
Alternatives
MiniMax-M3: common questions
- Is MiniMax-M3 worth it?
- Strong for its price on long, mixed-media inputs, held back by a licence that makes lawyers read twice. We rate it 3.9 out of 5.
- How much does MiniMax-M3 cost?
- Open weights under the MiniMax community licence. API per 1M tokens on prompts up to 512K: $0.30 input, $1.20 output, $0.06 cached; double above 512K; priority tier 1.5x.
- What are the downsides of MiniMax-M3?
- Licence needs attribution, a notice and permission above $20M revenue; mid-table on Artificial Analysis: 29, 23rd of 121 large open models; bF16 weights around 850 GB; aPI retention and training terms not found.
- Who is MiniMax-M3 best for?
- Cost-sensitive teams processing long documents or video who want an open model and can accept its licence terms.
- How much does MiniMax-M3 cost for a team or in production?
- 1M support replies a month at 2,000 input and 300 output tokens each is about $960 on the standard tier, less with cached system prompts. Self-hosting at BF16 needs roughly 850 GB of GPU memory before KV cache: an eight-GPU H200 or B200 node, or more than one H100 node.
- Is MiniMax-M3 safe for company data?
- Licence: commercial use needs a "Built with MiniMax M3" credit and a notice to MiniMax; products over $20 million in annual revenue need prior written permission; no military use. MiniMax's API retention and training terms were not published on the pages we could read.
- What are the best MiniMax-M3 alternatives?
- The closest alternatives we have reviewed are GLM-5.3, DeepSeek-V4.1-Flash, Kimi K3 and Gemini 3.8 Flash.
Answers checked Oct 23, 2026.