MiniMax M3.1 Flash
MiniMax · Vision-language · MoE 428B (A23B) · 1000k context · Released 23 October 2026
M3.1 Flash is a sparse mixture-of-experts model with, per its model card and configuration, around 428 billion total parameters (128 routed experts, four routed plus one shared active per token) and about 23 billion active for any given token. It accepts text and images and produces text, and supports a context window up to one million tokens, with usable length depending on memory and serving setup. The weights are published on Hugging Face in 8-bit (FP8) formats without a gate, and community quantisations for llama.cpp, Ollama and LM Studio appeared within days. At this scale it is a multi-GPU or large-unified-memory model rather than a workstation one. Note that Hugging Face's automatic metadata box shows a lower parameter figure (around 220B) that does not match the architecture and appears to be a miscount; the card's ~428B total is the accurate description.
Strengths
- A frontier-scale open-weight model that can be self-hosted with data kept in-house
- Natively multimodal, accepting images alongside text
- Very long one-million-token context window
- Sparse mixture-of-experts, so only about 23B of the ~428B parameters compute per token
- Community GGUF quantisations already available for llama.cpp, Ollama and LM Studio
Weaknesses
- Server-scale footprint, needing multiple GPUs or a large unified-memory machine rather than a consumer card
- Commercial use carries licence conditions (attribution, a one-time notice, and authorisation above 20 million US dollars in revenue)
- Capability figures so far are largely vendor-reported, with independent evaluation still limited
- Released primarily in FP8, so full-precision behaviour is not the published baseline
Hardware requirements
| Quantisation | Approx. VRAM | Notes |
|---|---|---|
| Q4 (approx) | ~215GB | A community 4-bit quantisation, roughly 215GB as a rule-of-thumb for the 428B total (verify against a specific GGUF). This fits a 256GB unified-memory machine or a multi-GPU server, so the model is not datacentre-only. Figure is approximate. |
| FP8 (as released) | ~430GB | The primary released format, roughly 430GB across many files, close to one byte per parameter. This is multi-GPU server territory, and actual needs rise with context and concurrency. Figure is approximate. |
What you'd need to run this
Licence
MiniMax Community License — read the licence
MiniMax M3.1 Flash: common questions
- What hardware do I need to run MiniMax M3.1 Flash?
- At its most compressed (Q4 (approx)) it needs roughly 215GB of VRAM. VRAM figures are approximate and depend on context length and settings.
- Is MiniMax M3.1 Flash free for commercial use?
- Commercial use is permitted, but with conditions. Open weights under the bespoke MiniMax Community License, not an OSI licence. Read the licence before relying on it at scale.
- Can I run MiniMax M3.1 Flash on Apple Silicon?
- It can run on Apple Silicon through general runtimes, but it is not specifically optimised for it.
- What is MiniMax M3.1 Flash's context window?
- MiniMax M3.1 Flash has a context window of 1,000,000 tokens, about 1000k.
Availability
Recommended for
- Organisations with multi-GPU or large unified-memory infrastructure wanting a frontier-scale open multimodal model
- Long-context and long-document work up to one million tokens
- Self-hosting where data must stay in-house and the licence's attribution and notice conditions can be met
Related models
Related guides
Catalogue entry last verified 27 October 2026. Specifications change; verify anything you are about to spend money on.