MiniMax has released M3.1 Flash, a lightweight coding model with a one-million-token context window and prices that start at $0.15 per million input tokens. It went live on October 19, after a quiet preview in late September and a mystery model that had been climbing leaderboards under the name Space Bunny. The pitch is easy to understand: big memory, fast replies and a low bill. The part nobody can fully judge yet is how good it actually is.
What M3.1 Flash is
M3.1 Flash is the fast, light member of MiniMax's M3 family. MiniMax describes it as a multimodal model built for coding, tool use, agent-style reasoning and long-context work. It accepts text, images and video as input and replies in text.
The standout feature is the context window of about one million tokens, enough to hold a very large codebase or a long stack of documents in a single session. That's a big jump for the company. The older M2 family topped out at 204,800 tokens, so the M3 generation is roughly five times larger.
The other headline feature is adjustable thinking depth. You can choose how hard the model reasons, trading answer quality against speed and cost. During the preview, the reported options were five levels: low, medium, high, extra high and max. That's useful in practice, because a quick bug fix doesn't need the same effort as a tricky design problem.
The Space Bunny story
This launch had an unusual build-up. Before the official name appeared, a stealth model called Space Bunny showed up on public testing platforms and drew attention for its speed and for how many tokens people were running through it. MiniMax has now confirmed that Space Bunny was an earlier version of M3.1 Flash, and it says the stealth model handled more than 200 trillion tokens.
That number is MiniMax's claim, and it measures usage, not quality. A model can be popular because it's free or fast without being the best. Still, a few hundred trillion tokens of real traffic is a serious stress test, and it suggests the company has been running this model at scale for a while.
The timeline looks like this: the preview launched inside MiniMax Code on September 27 with no price, no speed figure and no benchmark. On October 19, MiniMax's official account said the full model was coming, and the developer documentation already listed it without the "Preview" label. Pricing and the Space Bunny reveal followed.
Specs at a glance
- Developer: MiniMax
- Type: multimodal coding and agent model
- Context window: about 1 million tokens (1,048,576 on the largest deployment)
- Inputs: text, image, video
- Thinking control: adjustable depth, with five effort levels reported
- Tool calling: supported
- JSON output mode: not supported, according to one gateway listing
- Model ID: MiniMax-M3.1-Flash
- Released: October 19, 2026
What it costs
Pricing is the biggest reason people are paying attention. Gateway listings show the model starting at $0.15 per million input tokens and $0.60 per million output tokens, with cached input at $0.015. Those figures come from third-party gateways, so confirm them on MiniMax's own pricing page before you build a budget around them.
One report compares it with the larger M3 model. Token for token, M3.1 Flash costs about half as much as M3 below 512,000 tokens and about a quarter as much above that. In practice, that makes the long-context end the best value, which is exactly where most models get expensive.
The pricing also follows a pattern in the "Flash" tier from Chinese labs: a small, quick model aimed at everyday work, priced low enough that you can use it freely. The one-million-token window is the main argument that sets this one apart from its neighbors.
How you can use it
The preview was limited to MiniMax Code and to Token Plan subscriptions, so only people inside MiniMax's own tools could try it. With the full release, it's available through the API. MiniMax's documentation points to two endpoints: one that matches Anthropic's request format, which is the recommended one, and one that follows the OpenAI format. Both use the same model ID, so tools that already speak either format should be easy to point at it.
That compatibility matters for developers. If your coding assistant or editor lets you set a custom endpoint, switching to M3.1 Flash is mostly a settings change, not a rebuild. It's worth testing on a small task first, since support for specific features can differ between tools.
Open weights are coming
MiniMax has promised to release the model weights on October 23. That would follow the pattern set by the main M3 model, which was published with downloadable weights in early June. For developers, open weights mean the option to run the model on your own hardware, tune it for a specific job, and avoid depending on one company's servers.
Until the weights are actually posted, though, it's still a promise. Check the license when they arrive, because "open weights" can still come with limits on commercial use. Running a model with a one-million-token window also takes serious hardware, so self-hosting won't be realistic for everyone.
How fast is it?
Speed is the name of the game for a Flash model, and the early numbers are mixed. During the preview, one user reported decoding speeds around 90 to 110 tokens per second and called it slow for the quality. Since the full launch, a gateway's live traffic data has shown about 138 tokens per second, with roughly two seconds to the first token.
Those numbers come from different setups and different loads, so don't read too much into either one. Real speed depends on the provider, the thinking level you pick and how long your prompt is. If speed matters to you, run your own test with your own prompts.
What we still don't know
- Official benchmarks. As far as the reporting shows, MiniMax hasn't published a score for M3.1 Flash. The listings that mention benchmarks don't give numbers, and the figures for the larger M3 model don't carry over.
- The architecture. The documentation doesn't repeat the internal details. Treat anything borrowed from the M3 model as a guess until MiniMax confirms it for Flash.
- The weights' license. We'll know more once they're released.
- Real-world coding quality. Leaderboard attention and token counts don't show how it handles your project.
- Feature gaps. The missing JSON output mode could matter if your app depends on structured data.
Who should try it
M3.1 Flash looks best suited to everyday development work: bug fixes, small features, code review and the routine tasks that pile up in a queue. The huge context window also makes it a good fit for jobs that need a lot of material at once, like reading a whole repository or summarizing a long set of documents. The low price means you can run it often without worrying about the bill.
It's less clear that it's the right choice for the hardest problems, since MiniMax is positioning Flash as the quick, light option and not the flagship. If you need the strongest possible answers, the larger M3 model, which has a published score on a standard coding test, is the safer bet until Flash has numbers of its own.
The bottom line
M3.1 Flash is an interesting release mostly because of its price and its context window, and the Space Bunny story shows MiniMax tested it heavily before the launch. But the independent proof is missing. There are no official benchmarks, the weights haven't arrived, and the speed figures vary.
The smart move is to try it on a real task rather than trust a launch-week chart. Watch for the October 23 weights, the license terms, and independent coding tests over the next few weeks. Those will show whether this is a genuinely strong budget model or just a well-priced one.
