MiniMax Models: M3, M3.1 Flash and Where to Run Them

MiniMax's current open-weights models are MiniMax-M3 and MiniMax-M3.1-Flash: roughly 428 billion parameters in total with about 23 billion active per token, a 1M-token context, and text, image and video input. The weights sit on Hugging Face under a custom community licence. M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens on MiniMax's API. Checked on 31 October 2026.

The lineup in one table

The hub also holds an MXFP8 build of M3, the older M2.5, M2.1 and M2 text models, a music model, and H3, a video model with its own licence agreement. We stay with the text and multimodal language models here.

M3 came first, in June 2026 by the hub's creation date. M3.1 Flash was uploaded in October 2026 and its card describes it as building on M3 with sparse attention throughout the model, quantization-aware training and speculative decoding through a draft head. In other words, Flash is not a small model. It is the same class of size, tuned to be faster and cheaper to serve.

Maker, licence and what "open" means here

MiniMax is a Chinese AI company, and it publishes its weights on Hugging Face and its code on GitHub. The licence is the part to read carefully, because it is not a standard open-source licence. The hub tags it "other" and the file is named MiniMax Community License. Here is what the LICENSE file says in plain terms.

Two details are worth underlining. The ban on "any military purpose" is broader than most open licences, so a defence contractor should have a lawyer read it. And the fine-tuning clause means that a company that trains its own variant for a paid product is inside the commercial definition. The M3 file refers to "Built with MiniMax M3"; we did not check what attribution text the Flash file requires.

The term "open weights" describes this accurately. "Open source" in the sense of the Open Source Definition does not. We cover that distinction on the open source LLM page and in open versus closed models.

Context window and modalities

The M3 card says "native multimodal model with 1M context." It describes mixed-modality training from the first step and a new sparse attention operator, MiniMax Sparse Attention, which it credits with 9 times faster prefill and 15 times faster decode than M2 at 1M tokens. Those are MiniMax's claims in its own card, so read them as vendor-reported. The Flash card repeats the 1M figure and says usable length and throughput depend on GPU memory and serving setup. That caveat matters: a 1M window on paper is a different thing from a 1M window on your hardware.

Inputs are text, images and video. Output is text. M3 offers three reasoning modes through a thinking parameter: enabled, adaptive and disabled. The Flash card instead sets a reasoning effort in the chat template, with levels low, medium, high, xhigh and max. Recommended sampling on both is temperature 1.0 and top-p 0.95.

API prices

MiniMax's pay-as-you-go pricing page lists these figures in US dollars per million tokens.

A priority service tier is listed at 1.5 times the standard price. The page words the M3 discount as "permanent 50% off", so we show both numbers and treat the discounted one as today's price.

For a sense of scale, take a run of 10 million input tokens and 2 million output tokens. On M3.1 Flash at list that is $1.50 plus $1.20, or $2.70. On M3 at the discounted rate it is $3.00 plus $2.40, or $5.40. The same run on a $2 and $10 model such as Claude Sonnet 5.5 is $20 plus $20, or $40. We do not claim the cheap model gets the same results. We claim the list price gap is large enough that it is worth running your own tests.

Vendor-reported benchmarks

The M3.1 Flash card prints a comparison table. These are MiniMax's numbers, run by MiniMax, with the competitor figures as MiniMax reports them. We copy a subset and label it as vendor-reported. The card's evaluation notes name the test harness for several rows, such as Codex for Terminal Bench and mini-SWE-agent for DeepSWE, and say some competitor numbers come from official leaderboards or in-house reproduction.

Source for all rows: the MiniMax-M3.1-Flash model card on Hugging Face, vendor-reported. The card's model labels, "GPT-5.6 Sol" and "Opus 5", are MiniMax's and may not match the names the vendors use on their pricing pages today, so do not map them to the products in our price tables without checking.

What the table shows, read fairly: on repository tasks M3.1 Flash is in the same band as the other open models in the table, and the closed Opus column leads on several of the long agent benchmarks. On DeepSWE it is within about a point of DeepSeek and Opus in these numbers. A fair reading is "competitive, not dominant, and much cheaper." None of it replaces your own test.

Running it locally

The M3 card lists these serving stacks: SGLang, vLLM, Transformers, KTransformers, unsloth and AMD's ATOM. The Flash card lists SGLang, vLLM and TokenSpeed, with others in progress, and a multi-chip route through FlagOS. Download is a single command from the hub's CLI.

For the economics of renting GPUs for a model this size, see cloud versus self-hosted AI and the host pages for Runpod, Lambda and Modal. A model with 23 billion active parameters is cheaper per token than its size suggests, but all 428 billion must sit in memory.

Where it is served

MiniMax's own platform serves M3 and M3.1 Flash through its API and sells packages. The weights are free to download, so any host that runs open models can serve them in principle. We did not check which third-party hosts list an M3-series model today.. For hosts we cover, see AI hosting and Hugging Face Inference. Because the licence defines commercial API use as commercial use, a host that sells access to the model has its own obligations under the licence.

Coding agents that can use it

MiniMax ships its own agent, MiniMax Code, linked from both cards. For third-party tools, the Flash card's evaluation notes name the agent harnesses MiniMax used for its numbers, among them OpenCode, Codex and Claude Code, and say the model was scored in each. That tells us MiniMax tests against them. It does not tell us they officially support it. We did not find MiniMax's documented setup for them.. Any agent that takes an OpenAI-style base URL and a model name can in principle call it. See OpenCode and Cline for tools with flexible provider settings.

Privacy and data

Self-hosting keeps data on your machines. The hosted API is run by MiniMax and subject to its terms. We did not find the API's data retention, training use or server location terms on the pages we read.. If those terms matter to you, the usual answer is to run the weights yourself or on a host whose terms you accept.

Limits and gotchas

• The licence is not OSI open source. Commercial use needs attribution and a notice, and a large product needs written authorization.
• A military-use ban sits in the prohibited uses.
• Size. About 428B parameters means a multi-GPU server, even with 23B active.
• Context is a ceiling. Usable length depends on memory and the serving setup.
• Vendor scores only. The only benchmark table we found is MiniMax's own.
• Price tiers. M3 steps up at 512K input tokens, and priority costs 1.5 times standard.

Our take

MiniMax M3.1 Flash is the interesting one. At $0.15 and $0.60 it matches the off-peak price of DeepSeek's flash model on our price tables, it has a 1M window, and its weights are public. If you want a capable agent model and your volume is high, put it on your shortlist and run your own tasks. If your product is commercial, read the licence with a lawyer first. The attribution rule is easy to live with and the $20 million threshold will not bind a small team, but the military clause and the fine-tuning clause are real constraints.

If you need a cleaner licence, compare it with Apache and MIT options on the open source LLM page, including DeepSeek, Qwen and gpt-oss. Where it sits against closed models is in open versus closed models.

Who should pick it

Related pages

• Open source LLM for how open weights differ from open source
• Best open source LLM for coding for the shortlist
• Kimi models and GLM models for the other large Chinese open-weights families
• LLM API prices for a cross-vendor table