MiniMax M3
MiniMax's ~428B parameter natively multimodal Mixture-of-Experts model, handling text, images and video across up to 1M tokens of context, and a fit for a three-unit Faraday Scale cluster.
Model Story
MiniMax released MiniMax M3 on June 1, 2026 and published the weights on Hugging Face on June 11, 2026, together with an arXiv technical report. It is a Mixture-of-Experts model with roughly 428B total parameters and about 23B active per token.
M3 is natively multimodal: it takes text, images and video as input rather than bolting vision on afterwards. It supports up to 1M tokens of context, with 512K guaranteed on MiniMax's API and long-context requests billed at twice the standard rate.
That mix of long context and native multimodality is why Faraday Machines ships MiniMax M3 for long-context and multimodal agentic tasks: reviewing long video or image-heavy material, working through large document sets, and running multi-step agents that need to see as well as read.
Key Specifications
| Developer | MiniMax |
|---|---|
| Release Date | June 1, 2026; weights published June 11, 2026 |
| Architecture | Sparse Mixture-of-Experts (MoE) |
| Total Parameters | ~428 billion |
| Active Parameters | ~23 billion per token |
| Context Window | Up to 1M tokens (512K guaranteed on the API) |
| Multimodal | Native text, image and video input |
| Weights | Hugging Face, with an arXiv technical report |
| Size at 4-bit | Roughly 240 GB |
| Faraday Fit | Scale (3 Mac Studios, 384GB) |
API Pricing
MiniMax API pricing as of October 2026. Long-context requests are billed at 2× these rates, which adds up quickly for exactly the long-document and video workloads M3 is good at. On-premises deployment on Faraday Machines eliminates per-token costs entirely; you pay only for hardware amortization and electricity.
Benchmarks
All three scores are vendor-reported by MiniMax; we have not included an independent index score here. Treat vendor-reported SWE-bench Pro numbers with care: OpenAI's July 2026 audit estimated that about 30% of the public tasks are broken, and on Scale's standardized public set the top score is around 61.5%. Among open-weight models, MiniMax M3's 59.0% sits alongside other vendor-reported results such as GLM-5.2 (62.1%), Qwen3.8-27B (61.7%) and Kimi K2.6 (58.6%).
On-Premises Deployment
MiniMax M3 fits Faraday's Scale tier. At 4-bit quantization its weights are roughly 240 GB. That is technically under the 256GB of a two-unit Growth cluster, but it would leave almost no memory for long-context and multimodal working memory, which is the whole point of the model. On a three-unit Scale cluster (384GB of unified memory) there is real headroom for long prompts, images and video.
On Scale, MiniMax M3 pairs naturally with DeepSeek V4.1 Flash: M3 for multimodal and long-context agentic work, V4.1 Flash for agentic coding. Faraday helps you choose the right mix for your workload.
Frequently Asked Questions
Can I run MiniMax M3 on-premises?
Yes. MiniMax M3 runs on Faraday Machines' Scale tier: three Mac Studios with 384GB of unified memory. At 4-bit quantization its roughly 428B parameters need about 240 GB. That technically fits the 256GB Growth tier, but it would leave almost no memory for the long-context and multimodal work the model is built for, so Faraday Machines recommends the Scale tier.
How much does MiniMax M3 cost?
MiniMax's API charges $0.60 per 1M input tokens and $2.40 per 1M output tokens for MiniMax M3 as of October 2026, and long-context requests are billed at twice those rates. On-premises, Faraday Machines' Scale tier costs $29,999 US one-time, including hardware and 12 months of updates, maintenance and support, with no per-token charges.
Is MiniMax M3 multimodal?
Yes. MiniMax M3 is natively multimodal, taking text, images and video as input, and supports up to 1M tokens of context (512K guaranteed on MiniMax's API). MiniMax released it on June 1, 2026 and published the weights on Hugging Face on June 11, 2026. Faraday Machines ships it for long-context and multimodal agentic tasks such as reviewing video or image-heavy material.
How does MiniMax M3 compare to other open-weight models on coding?
MiniMax M3 is close to other leading open-weight models on vendor-reported coding scores. MiniMax reports 59.0% on SWE-bench Pro, alongside GLM-5.2 (62.1%), Qwen3.8-27B (61.7%) and Kimi K2.6 (58.6%). These are vendor-reported figures, and OpenAI's July 2026 audit estimated that about 30% of public SWE-bench Pro tasks are broken. MiniMax also reports 66.0 on Terminal-Bench 2.1.
References
[1] MiniMax. (2026). "MiniMax M3" model page. Available at: minimax.io
[2] Fireworks AI. (2026). "MiniMax M3 launch." June 2026. Available at: fireworks.ai
[3] Scale AI. (2026). "SWE-bench Pro (public set) leaderboard." Available at: labs.scale.com
[4] Morph. (2026). "SWE-bench Pro scores and caveats." Available at: morphllm.com
Explore Other Models
See how every model compares on price, benchmarks and hardware fit in the AI Model Comparison.
Run MiniMax M3 on Hardware You Own
A three-unit Faraday Scale cluster runs MiniMax M3 on-premises for long-context and multimodal work. No per-token charges, no 2× long-context surcharge, and full data sovereignty.
Get a Fixed-Cost Quote