MiniMax-M3.1-Flash: The Model behind Space Bunny, Explained
MiniMax Has Effectively Ended the Guessing Game by Tying Space Bunny to MiniMax-M3.1-Flash. This Guide Separates the Confirmed Specs: 1M Context, Native Image and Video Input, and Five Thinking Levels, from the Benchmark and Pricing Questions Still Open, so You Can Decide Whether to Test It Now.
On October 19, 2026, MiniMax's official account posted a two-line teaser: "Welcome back. Space Bunny. 🚀 MiniMax-M3.1-Flash is coming." For developers who followed Space Bunny, the post closed the loop. MiniMax has now tied its Space Bunny preview to MiniMax-M3.1-Flash, the next speed-focused model in its M series.
The reveal matters beyond internet speculation. The nickname now points to a named model with documented API access and a commercial path. It also sets an evaluation deadline: M3.1-Flash is "coming," not broadly launched, so now is the time to separate confirmed capability from hype.
| At a Glance | MiniMax-M3.1-Flash |
|---|---|
| Status | Announced October 19, 2026; API docs list the model |
| Context | 1,000,000 tokens |
| Inputs and output | Text, image, and video input; text output |
| Thinking | Always on; low, medium, high, xhigh, or max effort; max is the default |
| Architecture | 428B total parameters, about 23B active, sparse attention, native visual encoder |
| Speed | Approximately 150 tokens per second, according to MiniMax |
| Access | MiniMax Code, M Plan subscription key, Anthropic- and OpenAI-compatible endpoints |
What MiniMax Actually Confirmed
MiniMax's API documentation lists MiniMax-M3.1-Flash as its latest M-series model, with a 1,000,000-token context window and text, image, and video input. It calls the model a frontier multimodal coding model with tunable thinking depth.
Thinking is always on. Developers choose low, medium, high, xhigh, or max effort; the default is max. Unlike MiniMax-M3, there is no off switch: thinking: {"type": "disabled"} or effort: "none" returns a 400 error. The docs recommend lowering effort instead. Thinking arrives separately, so apps can show final answers without parsing <think> tags.
The API accepts Anthropic-compatible requests at https://api.minimax.io/anthropic and OpenAI-compatible requests at https://api.minimax.io/v1. Prompt caching is supported, which matters for long codebases and repeated agent turns.
MiniMax's official preview guide adds the architecture: a mixture of experts with 428 billion total parameters, about 23 billion active per token, sparse attention, and a native visual encoder. It quotes about 150 tokens per second. Those are vendor figures, not independent measurements.
What Changes from MiniMax-M3
| Capability | MiniMax-M3.1-Flash | MiniMax-M3 |
|---|---|---|
| Reasoning | Always enabled; five effort levels | Can be enabled or disabled |
| Context window | 1,000,000 tokens | 1,000,000 tokens |
| Multimodal input | Text, image, and video | Image and video |
| Prompt caching | Supported | Supported |
| Positioning | Fast everyday coding and agent work | Frontier open-weight model |
The practical change is control. M3.1-Flash asks how much reasoning to buy, not whether to reason at all. Max effort suits hard debugging and multi-step tool use; low effort keeps routine edits responsive. The tradeoff is that every request carries reasoning tokens, so output budgets must include them.
Where It Fits: Coding, Agents, and Visual Work
MiniMax positions M3.1-Flash for frequent coding, financial research, multimodal understanding, and frontend work. It targets tasks where the model must inspect material, use tools, and produce a working artifact, not just answer a question.
One case asks the model to diagnose a slow support-ticket interface, use browser performance tools, fix the bottleneck, and verify the result. Another turns a three-statement financial model and discounted cash flow into a reviewable workbook. A third compares images and explains visual differences. The most distinctive uses reference photos of the Qinian Hall to create an interactive Three.js reconstruction with orbit, zoom, and layered roof controls, with image understanding feeding code generation.
Benchmarks: The Evidence Is Workflow-Shaped
At publication, MiniMax has not released a standalone benchmark table for MiniMax-M3.1-Flash. The preview page offers documented workflows, not a scoreboard. That makes the model easier to understand and harder to rank.
The Flowdesk case is the most concrete performance evidence. MiniMax publishes a three-run comparison on the same 8,000-record dataset: 0.045 seconds after the fix versus 15.265 seconds before. The page calls it a constructed case that does not generalize: a workflow demonstration, not a speed promise.
MiniMax's official Flowdesk case compares three-run medians on the same 8,000-record dataset. The page calls it a constructed case, not a general speed guarantee. Source: MiniMax - M3.1-Flash-Preview documented case.
For context, MiniMax's M3 launch report lists 59.0 on SWE-Bench Pro and 66.0 on Terminal-Bench 2.1. Those are M3 scores under M3's settings, not M3.1-Flash results. Without a shared benchmark, a quality verdict is premature. Test your own repository-level bug fix, long-context retrieval task, image-to-code task, and multi-tool workflow, measured by completion rate and cost per finished task.
Pricing and Access: Subscription First
MiniMax's pay-as-you-go price table still lists MiniMax-M3, not M3.1-Flash. M3 costs $0.30 per million input tokens and $1.20 per million output tokens up to 512K input tokens, with higher rates above that. Do not assume M3.1-Flash inherits those numbers; as of October 19, 2026, it has no posted per-token rate.
The confirmed route is M Plan, with MiniMax Code and an API key for coding tools. Annual pricing is $220 for Go, $550 for Explore, and $1,320 for Build, each billed annually with two months free. All tiers list the M3.1 Flash Preview text model, image and audio generation, web search, and multimodal understanding. If you need metered billing today, M3 remains the safer choice; to evaluate M3.1-Flash, M Plan is the practical door.
Limitations and Who Should Wait
Three constraints matter. Thinking cannot be disabled, so latency-sensitive classification or extraction may be a poor fit. There is no published per-token price or standalone benchmark, which complicates procurement and capacity planning. And MiniMax has not announced open weights, so teams needing local deployment should stay with M3.
The model remains framed as a preview in several official surfaces, even though the API docs use the final model ID. Names, access rules, and rates can change before general availability. Build a fallback rather than hard-coding it into production.
Verdict: Test Now or Wait?
Test MiniMax-M3.1-Flash if you already pay for M Plan, run agent-heavy coding workflows, or need a fast model that turns screenshots and video frames into frontend code. Use low or medium effort for routine tasks and reserve max for debugging and multi-step tool use.
Wait if you need published benchmark scores, a per-token rate card, an SLA, or open weights. The teaser answers the identity question, not the procurement question. Space Bunny finally has a name; whether it earns a place in your stack is still an open test.
