Chinese company MiniMax put the open weights of M3.1 Flash online on October 23, 2026. The model is very cheap — only $0.15 for one million input tokens on its API. MiniMax says AI agent work costs 70% less with it. And now developers can download it and run it on their own computers.
What Happened
MiniMax launched M3.1 Flash on October 19, 2026. It is the faster and cheaper brother of the M3 model from June 2026.
Then on October 23 the bigger news came. MiniMax released the open weights on Hugging Face and ModelScope. Now any developer can download the model and run it on his own machine. No need to depend on MiniMax servers.
Tech media covered it a lot. SiliconANGLE reported the launch on October 19 and Kie.ai confirmed the weights release on October 23.
The Mystery of “Space Bunny”
Before launch, a strange model with the name “Space Bunny” showed up on OpenRouter. Nobody knew who made it. It handled more than 200 trillion tokens and stayed on top of the OpenRouter charts for 10 days.
Now MiniMax told the truth. Space Bunny was an early test version of M3.1 Flash. The company tested it in public with no name and then launched it as M3.1 Flash.
Key Details of the Release
Here are the main points, in simple words:
Very low price. API price is $0.15 for one million input tokens and $0.60 for one million output tokens. Cached input is only $0.015. And the price stays same even for very long prompts.
70% cheaper agent work (company claim). MiniMax says AI agents cost 70% less to run than on the older M3. It also says the model is nearly three times faster, with output above 200 tokens per second. But remember, these numbers come from MiniMax itself. No outside lab has checked the speed claim till now.
Big model with smart design. M3.1 Flash has 428 billion parameters in total, but it uses only around 23 billion for each token. This design is called mixture-of-experts. It sends each task to a small group of expert parts. That is why it stays fast and cheap.
One million token context. The model can read up to one million tokens in one prompt. That means one full book or one big codebase. It takes text, images, and video as input.
Thinking is always on. This model always “thinks” before it answers. You can pick from five effort levels: Low, Medium, High, Xhigh, and Max. Max is the default. You cannot switch thinking off. If you try, the API gives an error.
Now open weights. On October 23, MiniMax put the model files on Hugging Face and ModelScope. The files are around 252 GB. Early users say they are running it on a 4x NVIDIA DGX Spark setup.
License has some rules. The model uses the “MiniMax Community License”. It is not as free as MIT. If your company earns less than $20 million in a year, just send a notification email. And show a “Built with MiniMax M3.1-Flash” credit in your product. If you earn more than that, you need written permission. And military use is not allowed.
How It Compares
Artificial Analysis is one independent test site, and it gave M3.1 Flash a score of 42 on its Intelligence Index. The older M3 got 29. So the jump is real, and an outside party confirmed it.
The closest rival is DeepSeek V4.1 Flash 1017. Both cost $0.15 per million input tokens in off-peak hours. We covered the DeepSeek V4.1 Flash 1017 release on this site.
On one automation test MiniMax wins, 62.2 against 60.5. On the DeepSWE coding test DeepSeek wins, 77.0 against 73.7. So use MiniMax for office-style agent work and DeepSeek for hard coding jobs.
If you like open models, also read our post on Mistral Large 4, another open-weights release.
Why It Matters for Freelancers
API cost eats the profit of AI freelance work silently. If you sell chatbots or automation, every API call takes money from your pocket. A model this cheap changes the full math.
Open weights give you a second win. You can host the model on your own server, and your client data never leaves your hands. That is big for clients with strict privacy rules.
The one-million-token context is the third win. You can feed long documents, full websites, or big spreadsheets in one go. Less calls means less cost and faster work.
Why It Matters for Students
Students who build AI projects get a new budget option. This API price is one of the lowest in the market, and small experiments cost almost nothing.
The open weights also help in learning. You can see how a real mixture-of-experts model is built. But be honest about hardware. The files are 252 GB, and you need serious GPUs to run them. A normal laptop cannot do it.
One warning for students. Artificial Analysis found a 93% hallucination rate. This model almost never says “I don’t know” and it guesses instead. So always double-check facts for homework and exams.
What to Watch Out For
No release is perfect. Some honest points:
Most numbers come from MiniMax. The 70% cost cut and the speed claims are company numbers. Only the Intelligence Index score of 42 comes from an independent test. So wait for more outside reviews before you trust every claim.
The license is not fully open. The Community License has revenue limits. Read it first before you build a business on this model. It is not MIT like some other open models.
Privacy needs one thought. MiniMax is a Chinese company. If you handle private client data on their API, read the privacy policy first. Self-hosting the weights solves this problem.
How to Try It
You can start like this:
- API users can sign up on the MiniMax platform. Price is pay-as-you-go.
- The model is also inside the MiniMax Code app and the M Plan subscription.
- Developers can download the weights from Hugging Face or ModelScope. But you need strong hardware with lots of GPU memory.
The Bottom Line
Two things make this release big. The price is tiny, and the model is now downloadable.
MiniMax is fighting a price war with DeepSeek and the big US labs. For freelancers and students, this war means more power for less money. Just remember the caveats. Check the benchmarks and read the license. And never trust a model that never says “I don’t know”.
Sources
- SiliconANGLE: MiniMax launches M3.1 Flash model, says it cuts AI agent costs by 70%
- Kie.ai: MiniMax M3.1 Flash Release: Benchmarks, Pricing, Limits (weights release confirmed Oct 23)
- MarkTechPost: MiniMax Just Launched M3.1 Flash: 428B MoE Agentic Model at $0.15/M Input