Which PC for MiniMax-M3.1-Flash? Memory, machine and license
Updated: November 1, 2026
What PC for MiniMax-M3.1-Flash?
A multi-GPU server, not a PC.
MiniMax-M3.1-Flash has 428 billion parameters. In 4-bit, it requires about 247 GB of GPU memory. No single card is enough, nor even two RTX PRO 6000s: you need at least three, so the Threadripper workstation with 3 or 4 cards, or the 8-GPU server.
This model appeared on Hugging Face on October 17, 2026. It is a multimodal mixture-of-experts model: it reads text, images, and videos, and responds in text. MiniMax intends it for coding, agents, computer use, and cybersecurity. Here is what is needed to run it at home, calculated, and what we cannot yet say.
The memory it needs
An expert model loads entirely into memory, even if only a part works on each word. You must therefore count the 428 billion parameters, not the 23 billion active ones. Our calculation adds the context cache (32,768 tokens) and an operating margin.
| Weight precision | GPU memory | Machine from our range |
|---|---|---|
| Q4 (4 bits) | 247 GB | CoreIA Threadripper PRO 4 GPU with 3 or 4 RTX PRO 6000 (288 or 384 GB), or EPYC 8 GPU AI Server, base configuration (4 cards, 384 GB) |
| Q8 (8 bits) | 429 GB | EPYC 8 GPU AI Server, with optional cards (up to 768 GB) |
| FP16 (16 bits) | 802 GB | No machine in our range |
Estimates based on the size announced on the model card, without measurement on our machines. The published repository weighs 235 GB, or about 4.7 bits per parameter: this is consistent with the Q4 line.
See the EPYC 8 GPU AI Server
See the CoreIA Threadripper PRO 4 GPU
Which machine
| Machine | GPU memory | Price excl. VAT | MiniMax-M3.1-Flash |
|---|---|---|---|
| CoreIA 2x RTX PRO 6000 | 192 GB | €40,423 | No, it lacks at least 55 GB in Q4 |
| CoreIA Threadripper PRO 4 GPU | 96 GB base, 384 GB with 4 cards |
€27,206 with 1 card, the additional cards are added |
Yes in Q4 with 3 or 4 RTX PRO 6000 (Max-Q versions beyond 2 cards); not in Q8 |
| EPYC 8 GPU AI Server | 384 GB base, 768 GB maximum | €171,791 | Yes in base Q4; Q8 with optional cards |
It is a high-end workstation or server investment, not a desktop. It is justified if you want a model of this level at home, without sending your data to a provider. The server comes with the software installed and tested.
Its level, according to MiniMax
The model card compares MiniMax-M3.1-Flash to other large models, not to Qwen3.8-27B or Gemma 4 12B: we therefore cannot place it against them with published figures. Here are a few lines from this table, against the open models mentioned.
| Test (scores published by MiniMax) | MiniMax M3.1 Flash | DeepSeek V4.1 Flash-1017 | Kimi K3 | GLM-5.3 |
|---|---|---|---|---|
| DeepSWE v1.1 (code) | 73.7 | 77.0 | 67.5 | 66.9 |
| Terminal Bench 4.0 (terminal agent) | 38.1 | 42.3 | 12.6 | 37.9 |
| NL2Repo-Bench (code of an entire repository) | 61.2 | 68.2 | 58.0 | 58.0 |
| HLE (without tools) | 43.4 | 42.6 | 43.5 | 42.3 |
| BabyVision (with tools) | 91.0 | 92.5 | 85.7 | not published |
Scores given by MiniMax, measured with its own settings: an order of magnitude, not an independent ranking.
The license, before any business use
Free except for commercial use
The MiniMax community license allows free non-commercial use.
Commercial use: a mention and a notification
You must display "Built with MiniMax M3.1-Flash" on your site, your interface, or your documentation, and notify MiniMax by email. Beyond 20 million dollars of annual revenue, prior written authorization is required.
A fine-tuned model also counts
Deploying a reworked version (fine-tuning, post-training) for commercial purposes falls under commercial use. Some uses are prohibited, notably military.
Summary of the license text published with the model, which alone is authoritative. This is not legal advice.
Do you really need a model of this size?
Often, no. For an internal assistant or searching your documents, a model of 20 to 30 billion parameters fits on a single card, for a much lower budget. A model of this size is justified for agents that code or drive a computer, when quality takes precedence over cost.
| Machine | GPU memory | Price excl. VAT |
|---|---|---|
| CoreIA RTX 5090 | 32 GB | €8,576 |
| CoreIA Radeon AI PRO R9700 | 32 GB | €3,469 |
Which model for your use?
Tell us what you want to run. We tell you which machine, and if a smaller model is enough. Free quote, no commitment.
Frequently asked questions
What GPU memory is needed for MiniMax-M3.1-Flash?
About 247 GB in 4 bits, 429 GB in 8 bits, and 802 GB in 16 bits, with a context of 32,768 tokens. These are estimates based on the announced 428 billion parameters. The published repository weighs 235 GB.
Can it be run on a single graphics card?
No. Even the RTX PRO 6000, with 96 GB, is far from enough. Two cards (192 GB) are not enough either in 4 bits. You need three: our CoreIA Threadripper PRO 4 GPU accepts 3 or 4 RTX PRO 6000 (Max-Q versions), and the EPYC 8 GPU AI Server has four by default.
Why does the repository display about 220 billion parameters?
Its weights are stored in a compressed format, which reduces the number of counted values. The model card announces 428 billion parameters, including 23 billion active for each word.
Which engine to use to serve it?
MiniMax recommends SGLang and vLLM, with recipes published on the model card, as well as TokenSpeed.
Can it be used in a company?
Yes, under conditions: the community license requires displaying "Built with MiniMax M3.1-Flash" and notifying MiniMax, with written authorization beyond 20 million dollars of annual revenue. Read the full text of the license.
What speed can be expected?
We have not measured this model on our machines and therefore do not announce any speed. It depends on the GPU memory and the engine setting.
Sources: MiniMax, MiniMax-M3.1-Flash card on Hugging Face; MiniMax, community license; DeepSeek, DeepSeek-V4.1-Flash-1017 card. Retrieved November 1, 2026: the memory figures are estimates from Radiance Systems.
