Updated Oct 22, 2026: MiniMax has since released the M3.1 Flash weights on Hugging Face, on the evening of Oct 22 (about 8:59 PM Toronto time), under the MiniMax Community License. Details below.
"MiniMax M3.1" went from almost no posts on X to about 2,100 on Oct 19 alone. That's the day MiniMax revealed what "Space Bunny" was. The draw is simple: near-frontier coding scores at a price that rounds to zero for most side projects.
Space Bunny, revealed
Before launch, an anonymous model called Space Bunny was handling huge volumes of traffic. On Oct 19, 2026, MiniMax said it was an early version of M3.1 Flash, and that it processed more than 200 trillion tokens while anonymous.
The price
| Token type | Price per 1M tokens |
|---|---|
| Input | $0.15 |
| Output | $0.60 |
| Cached input | $0.015 |
MiniMax says agentic usage costs 70% less than its previous model, and online inference is "nearly three times faster". It quotes output speeds of 200+ tokens per second, with a median of 170+.
For scale, OpenAI's GPT-6 Astra is $10 input and $50 output. Our API pricing comparison puts M3.1 Flash next to Claude and Gemini.
The benchmarks (MiniMax's numbers)
| Benchmark | What it tests | M3.1 Flash |
|---|---|---|
| DeepSWE v1.1 | Real-world software tasks | 73.7% |
| OSWorld V2 | Using a computer across apps | 61.7 |
| Agents' Last Exam | Professional computer-use tasks | 29.6 |
| CyberGym | Finding and validating vulnerabilities | 82% |
MiniMax says it ranks first among Chinese models on OSWorld V2 and Agents' Last Exam. These are vendor-run scores. They're a reason to test the model, not proof it beats the one you use now.
Where to use it
- MiniMax API (platform.minimax.io)
- M Plan, MiniMax's subscription
- MiniMax Code, its desktop coding app
- OpenRouter, as
minimax/minimax-m3.1-flash. The OpenRouter page lists a 1M-token context and up to 524,288 output tokens. The free launch period ended on Oct 22.
Most coding agents, including OpenCode and Claude Code-style tools, can point at an OpenAI-compatible endpoint. Our OpenCode provider guide shows the config pattern.
The open weights are out
Update, Oct 22: MiniMax's blog said on Oct 19 that "model weights will be released later", and outlets reported an Oct 23 date citing a MiniMax post on X. The weights came a little early. The MiniMax-M3.1-Flash repository went public on Hugging Face on the evening of Oct 22, 2026. Its latest update is stamped about 8:59 PM Toronto time. When we checked at 8:28 PM, it wasn't there yet.
What the model card says:
- Size: about 428B total parameters, with 23B activated per token (a mixture-of-experts model).
- Format: MXFP8 weights. The repository's files add up to roughly 252 GB, per Hugging Face's file listing.
- Inputs: text, images and video, with a context window of up to 1M tokens.
- Serving: MiniMax points to SGLang and vLLM for local deployment.
At that size, this is a multi-GPU server model, not a laptop model. Our VRAM guide for local LLMs and the local AI hardware sizing sheet help you tell what it would take.
The license: free for non-commercial use, conditions for commercial use
The weights ship under the MiniMax Community License, not MIT or Apache 2.0. In plain terms:
- Non-commercial use is free. You can use, copy, modify, merge, publish, distribute and sublicense the model for non-commercial purposes, as long as you keep MiniMax's copyright and permission notice.
- Commercial use is allowed with conditions. You must prominently display "Built with MiniMax M3.1-Flash" on a related website, user interface, blog post, about page or product documentation. Then:
- If the product or service makes more than $20 million a year in revenue, you need separate, prior written authorization from MiniMax.
- Below that, you send MiniMax a one-time notice. The license explains how to send it.
- "Commercial use" is defined broadly. It covers paid products or services that rely on the model, commercial use of APIs that serve it (hosted or cloud included), and commercial deployment of fine-tuned or otherwise modified versions.
- Some uses are banned outright. An appendix rules out illegal content, any military purpose, harming minors, false information meant to cause harm, and discrimination or hate speech.
- No warranty. The weights come "as is".
If you only call M3.1 Flash through MiniMax's API or OpenRouter, the API provider's terms apply instead. The license matters when you download and run the weights yourself.
Who should try it
- Cost-sensitive agent builders. At $0.60 per million output tokens, long agent loops get cheap.
- Anyone on a $20/month budget who keeps hitting usage caps on bigger models.
- Teams who want a fallback model for simple steps. Route easy calls to Flash and save the expensive model for hard ones.
Before you switch, run your own tasks. A cheap model that needs three retries isn't cheap. MiniMax also says a larger Pro model "will also arrive later".
