A model that spent months answering queries under a cartoon codename now has a public download and a price list. MiniMax put the M3.1 Flash weights online this week, and the license has a revenue line in it.
MiniMax put the M3.1 Flash weights on Hugging Face on October 23
The model itself arrived on October 19. The weights took four more days. MiniMax announced the upload on X at 01:06 UTC on October 23, pointing to Hugging Face and ModelScope, and the model card went live with the full configuration.
That gap matters for anyone who read the launch post. It said weights would be released later. They now are, which turns M3.1 Flash from an API product into something you can download, inspect and run yourself.
The download is large. Third-party write-ups put the repository at about 252 GB. Budget the disk before you budget the afternoon.
Inside M3.1 Flash's 428B parameter, 23B active design
The model card describes a native multimodal mixture-of-experts system: roughly 428 billion parameters in total, with about 23 billion active for each token. Text, images and video go in. Text comes out.
That ratio is the whole cost story. A dense 428B model would need every parameter to fire on every token. A sparse model wakes a slice of the network instead, so you pay for 23B of compute while carrying 428B of knowledge.
Two other pieces do the heavy lifting. MiniMax Sparse Attention runs through the entire model and keeps long contexts from turning into a quadratic bill, which is the difference between a 1M-token window you can afford and one you can only advertise. A DSpark speculative decoding draft head guesses the next tokens ahead of the main model, then lets it verify the guess in one pass. Training also used quantization-aware training, which bakes low-precision math into the model instead of bolting it on afterward.
MiniMax says the result is online inference about three times faster than the previous generation, with everyday output above 200 tokens per second and a median of 170 or more.
Space Bunny was M3.1 Flash in disguise
Before the name existed in public, the model ran on OpenRouter as "Space Bunny". No vendor label, no model card, just an anonymous endpoint with unusual speed.
MiniMax now says Space Bunny processed more than 200 trillion tokens while it operated without a name, and that it was an early version of M3.1 Flash.
Why mention it? Because anonymous testing is how labs collect real traffic before a launch, and 200 trillion tokens is a lot of real traffic. It's also a claim you can't check from outside. The company is the only source for that number.
M3.1 Flash benchmarks, all reported by MiniMax
The numbers below come from the launch post and the model card. No outside lab has reproduced them.
DeepSWE v1.1, which tests whether a model can read a codebase, add a feature and validate the fix, gives 73.7%. OSWorld V2, a computer-use test built around real desktop tasks, gives 61.7. Agents' Last Exam, which scores work with economic value attached, gives 29.6.
MiniMax says M3.1 Flash ranks first among Chinese models on both OSWorld V2 and Agents' Last Exam. That's a narrower claim than it looks, and a fair one to make.
Fair framing. Not proof.
On the security side, CyberGym comes in at 82.2, and BankerToolBench at 76.2 for multi-step financial workflows. MiniMax also says the model sits in the top group on AutomationBench-AA, without publishing a rank.
Read the set as a pattern rather than a leaderboard. MiniMax is claiming parity with top models on agent tasks at a much lower cost per task. The cost half is checkable today. The parity half needs independent runs.
What the MiniMax Community License requires
The weights ship under the MiniMax Community License. Hugging Face tags it as "other", which is accurate and unhelpful in equal measure.
Reported terms are worth knowing before you build a product on it. The license is free for non-commercial use, commercial products are expected to carry a "Built with MiniMax M3" attribution, and companies above roughly $20 million in annual revenue need separate written authorization from MiniMax.
That last clause is the one to read twice. A hobby project and a funded startup are working under very different obligations here, and both of them pulled the same files from the same repository. These terms come from third-party readings of the license text, so check the license in the repository yourself if the stakes are real.
M3.1 Flash pricing and what it does to agent budgets
Flat rates: $0.15 per million input tokens, $0.60 per million output, $0.015 for cached input. No long-context surcharge, no tier ladder.
Put that next to the 1M-token context and the agent story writes itself. An agent that reads a large document, calls six tools and revises its answer can burn millions of tokens in a single task. At these rates, the same run costs cents instead of dollars.
Is cheaper always better? No. A cheap model that loops three extra times can cost more than an expensive one that gets it right on the first pass. That's why the benchmark claims and the price list have to be read together, and why task-level cost matters more than token price.
MiniMax puts the overall usage cost at 70% below its predecessor. That figure assumes the model finishes in fewer steps, which is exactly the part outside testers haven't measured yet.
Where to run M3.1 Flash, and what to watch
The model is live on the MiniMax API, in M Plan and inside MiniMax Code. The weights are on Hugging Face and ModelScope for anyone with the disk space and the patience.
Three things to watch over the next month. Independent agent benchmarks, since that's where the cost advantage either holds or collapses. Real-world throughput on your own hardware, because 200 tokens per second on a vendor cluster says little about a single GPU. And license interpretation, since a revenue threshold that's vague in practice tends to get tested by the first company that crosses it.
If you're building an agent that reads a lot and reasons a little, this is one of the cheapest ways to do it right now. If you're building one that has to be right the first time, wait for someone outside MiniMax to run the tests.
