MiniMax releases open weights for M3.1 Flash, under a license with strings attached for commercial use
Announced · Published · By Contractor AI Shop
MiniMax has published downloadable weights on Hugging Face for M3.1 Flash, the coding and agent model it put on its API Oct 19. The model card describes a multimodal mixture-of-experts model with about 428B total and 23B active parameters, a 1M-token context and the MiniMax Community License.
Key facts
- This is a follow-up to our Oct 20 brief on MiniMax's M3.1 Flash API launch. MiniMax (Hugging Face model card)
- The weights went public on Hugging Face on Oct 22, Pacific time (Oct 23 in Beijing), and were up by about 6:10 PM PT, when the first Chinese press reports ran. Yicai / 第一财经 (Liu Xiaojie), Chinese-language
- The model card lists approximately 428B total parameters with 23B activated per token, text, image and video input, and a context window of up to 1M tokens. MiniMax (Hugging Face model card)
- Under the MiniMax Community License, commercial use requires prominently displaying "Built with MiniMax M3.1-Flash", and products or services with more than $20M in yearly revenue need prior written authorization from MiniMax. MiniMax (Hugging Face model card)
- The independent coverage so far is Chinese-language. IT Home (Wang Miao) reports the 428B total / 23B active figures and lists day-one partners including NVIDIA, AMD, Together AI and Fireworks. IT Home / IT之家 (Wang Miao), Chinese-language
- Yicai (Liu Xiaojie), also in Chinese, reports MiniMax's claim that users processed more than 100 trillion tokens in the 72 hours after the Oct 19 launch. Yicai / 第一财经 (Liu Xiaojie), Chinese-language
Our take: what it means for contractors
This is a big open model, not a Mac-Studio-class download. At about 428B parameters it needs multi-GPU servers (the card recommends SGLang, vLLM or TokenSpeed). For contractors the useful part is the license: commercial use requires a "Built with MiniMax M3.1-Flash" notice, and companies with more than $20M in yearly revenue need MiniMax's written permission first. Most readers should keep using the hosted API. The high hallucination rate from our Oct 20 brief still applies.
What we don't know yet
- Whether Ollama, LM Studio or MLX builds that fit a single high-memory Mac will appear
- Independent benchmarks of the open weights
- When MiniMax will update its Oct 19 blog, which still says weights come "later"
Sources
- MiniMax (Hugging Face model card) · MiniMaxAI/MiniMax-M3.1-Flash · 2026-10-22 · Primary
- IT Home / IT之家 (Wang Miao), Chinese-language · MiniMax M3.1 Flash 开源模型权重，总参数 428B · 2026-10-23 · Secondary
- Yicai / 第一财经 (Liu Xiaojie), Chinese-language · MiniMax M3.1 Flash上线三天处理超100万亿Token，今日正式开源 · 2026-10-23 · Secondary