othermultimodalRuns locally428B (card)
MiniMax M3.1 Flash: Specs, License & How to Run
Open-weight multimodal MoE (card 428B/23B, 1M context) under the MiniMax Community License.
- License
- other
- Parameters
- 428B (card) (active 23B (card))
- HF likes
- 259 (2026-10-24)
- First seen
- 2026-10-17
TL;DR. MiniMax M3.1 Flash is a multimodal model from MiniMax. Parameters: 428B (card) (active 23B (card)).License: other. It is listed under ChatGPT, Closed LLM APIs.
| Developer | MiniMax |
|---|---|
| Type | multimodal |
| Architecture | Native multimodal MoE (card model_type minimax_m3_vl; architecture MiniMaxM3SparseForConditionalGeneration). MiniMax Sparse Attention (MSA) throughout; QAT NVFP4 on routed experts; DSpark speculative-decoding draft head. Hugging Face safetensors.parameters sum to 220,490,854,016. |
| Parameters | 428B (card) · active 23B (card) |
| Modalities | text, image, video |
| Context | 1,000,000 tokens |
| License | other — LICENSE file title is MINIMAX COMMUNITY LICENSE (copyright 2026 MiniMax). Free for non-commercial use with the copyright notice. Commercial Use requires a prominent “Built with MiniMax M3.1-Flash” attribution. If yearly revenue exceeds USD 20 million, prior written MiniMax authorization is required; otherwise a one-time notice. Contact details and a prohibited-uses appendix are on that LICENSE file. Hugging Face cardData is license other / license_name minimax-community. This is not MIT or Apache-2.0. Not legal advice. |
| Hardware | Not yet verified — Card usedStorage is 252,429,151,663 bytes (~235 GiB). NVIDIA NIM card names Blackwell B200 / B300 / GB200 / GB300 for the hosted trial. Peak VRAM is not independently measured. Card points at SGLang, vLLM, and TokenSpeed recipes. (not independently tested) |
| Hugging Face likes | 259 (as of 2026-10-24) |
| Model card | https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash |
| First seen | 2026-10-17 |
MiniMax-M3.1-Flash is a native multimodal Mixture-of-Experts model from MiniMax. Official hosts we opened on 2026-10-24: the ungated card MiniMaxAI/MiniMax-M3.1-Flash (created 2026-10-17T06:37:50Z; last modified 2026-10-23T03:37:20Z; 259 likes / 1,096 downloads), the MINIMAX COMMUNITY LICENSE, the blog MiniMax M3.1 Flash (dated 2026-10-19), and NVIDIA’s NIM model card.
This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT. Head-to-head facts versus the MIT 1017 checkpoint already on this site: MiniMax M3.1 Flash vs DeepSeek-V4.1-Flash.
The card positions the model for coding, agentic workflows, computer use, and cybersecurity. Inputs are text, image, and video; output is text. Hugging Face pipeline_tag is image-text-to-text. Those sentences stay vendor.
What we verified (2026-10-24)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash (not gated) | HF API |
| Created / last modified | 2026-10-17T06:37:50Z / 2026-10-23T03:37:20Z | same |
| Likes / downloads | 259 / 1,096 | same |
| License tag + file | other / minimax-community · MINIMAX COMMUNITY LICENSE | HF cardData + LICENSE |
| Pipeline tag | image-text-to-text | HF API |
| Architecture | minimax_m3_vl (multimodal MoE) | card + HF config |
| Params (card) | ~428B total / 23B activated | card |
| safetensors.parameters | 220,490,854,016 | HF API |
| usedStorage | 252,429,151,663 bytes (~235 GiB) | same |
| Context | up to 1,000,000 tokens | card |
| Sampling (card) | temperature 1.0, top_p 0.95 | card |
| Reasoning effort | low / medium / high / xhigh / max | card |
Hugging Face’s UI can show a ~220B “model size” line from the safetensors sum. We keep the README’s 428B / 23B quote and do not invent a reconciliation — same pattern as DeepSeek-V4.1-Flash.
The 19 October 2026 blog still says weights “will be released later.” The card we opened on 2026-10-24 already lists 48 safetensor shards plus a dspark/ draft head. Treat the blog sentence as older product copy, not a 404.
How to run (from the card, not tested)
Card download:
hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-FlashRecommended serving frameworks copied from the README: SGLang, vLLM, TokenSpeed. Additional integrations listed as in progress: Transformers, KTransformers, unsloth, ATOM. Recommended sampling: temperature 1.0, top_p 0.95. Reasoning uses reasoning_effort in the chat template.
Hosted platform.minimax.io and MiniMax Agent are API / product SKUs. Flat API prices printed on the blog ($0.15 / $0.60 / $0.015 per 1M tokens) are not a statement about the weight files. forkalt has not served the shards.
License
LICENSE is the MiniMax Community License, not MIT and not Apache-2.0.
Copied conditions (paraphrased; read the file):
- Non-commercial use is granted with the copyright notice.
- Commercial Use requires a prominent “Built with MiniMax M3.1-Flash” attribution on a related website, UI, post, about page, or docs.
- If products or services generate more than USD 20 million yearly revenue, you need prior written authorization. Otherwise a one-time notice. Addresses and subject lines are printed on the LICENSE file.
- An appendix lists prohibited uses (illegal content, military purpose, harm to minors, and other listed categories).
Not legal advice.
What we did not invent
- VRAM, tokens/s (including the blog’s “200+ TPS” line), and “generational leap” stay vendor.
- Card bench rows stay self-reported.
- We do not rank MiniMax against ChatGPT or DeepSeek-V4.1-Flash.
Related: MiniMax M3.1 Flash vs DeepSeek-V4.1-Flash, DeepSeek-V4.1-Flash, Kimi K3, Reflection Beam, Raon-1-mini, vLLM, DeepInfer vs vLLM. Self-hosted UIs: Open WebUI, LibreChat, Jan. Pixel-editor traffic on this site still lands on PhotoCraft vs GIMP — a different job.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| DeepSWE v1.1 (card) | 73.7 (self-reported) | https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash |
| Automation-Bench v1.0.6 (card) | 62.8 (self-reported) | https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash |
| OSWorld 2.0 (card) | 61.7 (self-reported) | https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash |
| BankerToolBench (card) | 76.2 (self-reported) | https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash |
Deploy options
- local: Hugging Face weights
- local: vLLM recipe (card)
- hosted: MiniMax API (separate SKU)
Comparisons
FAQ
Are the MiniMax M3.1 Flash weights actually public?
Yes on the host we opened. MiniMaxAI/MiniMax-M3.1-Flash is ungated (created 2026-10-17T06:37:50Z; last modified 2026-10-23T03:37:20Z; 259 likes / 1,096 downloads in the 2026-10-24 API check). The 19 October 2026 MiniMax blog still says “Model weights will be released later”; the Hugging Face card we opened already hosts the weight shards. Not the same artifact as the MiniMax API SKU.
Is this MIT or Apache-2.0?
No. The LICENSE file is the MiniMax Community License. Commercial Use needs the “Built with MiniMax M3.1-Flash” attribution. Yearly revenue over USD 20 million needs prior written authorization; otherwise a one-time notice. Card tag is other / minimax-community. Not legal advice.
Does it replace ChatGPT or beat DeepSeek-V4.1-Flash-1017?
No sourced bake-off on this site. The card prints a self-reported table that names DeepSeek V4.1 Flash-1017 and other vendors. Those rows stay vendor. Head-to-head facts we can copy: MiniMax M3.1 Flash vs DeepSeek-V4.1-Flash.
How big is it?
Card: approximately 428B total / 23B activated per token, 1M context. Hugging Face safetensors.parameters sum to 220,490,854,016. We keep the README 428B/23B quote and do not invent a reconciliation. usedStorage 252,429,151,663 bytes.
Sources & verification
- https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash
- https://huggingface.co/MiniMaxAI/MiniMax-M3.1-Flash/blob/main/LICENSE
- https://www.minimax.io/blog/minimax-m31f
- https://build.nvidia.com/minimaxai/minimax-m3.1-flash/modelcard
- https://github.com/vllm-project/vllm
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.