What Is MiniMax M3.1-Flash-Preview? Space Bunny Alpha (2026)
The short answer
MiniMax M3.1-Flash-Preview is a 1M-context coding model that MiniMax quietly enabled inside its MiniMax Code agent on September 27, 2026 — no model card, no benchmarks, no API listing, no price. Its public facts fit in one line: up to 1M tokens of context, five reasoning-effort tiers (low, medium, high, xhigh, max), and a positioning of “fast, reliable, everyday development.” It is very likely the model that ran anonymously on OpenRouter as Space Bunny Alpha at $0 four days earlier, based on community tokenizer fingerprinting that matched MiniMax 24/24 and 50/50.
What MiniMax actually said
The entire announcement came from the MiniMax_Agent account on X: the model “debuts today on MiniMax Code” and is “built for everyday development, fast, reliable, and ready for real work, from quick bug fixes to full features.” That is it. You cannot find M3.1-Flash-Preview on OpenRouter, in the MiniMax API docs, or on a pricing page as of September 28, 2026.
Two things are documented inside MiniMax Code:
- Context up to 1M tokens.
- Five effort levels: low, medium, high, xhigh, max. The fifth tier, max, is new. More effort means longer thinking, more output tokens and higher latency — the standard reasoning-model trade — but a five-step dial rather than three suggests MiniMax is tuning for developers who want to trade cost against task difficulty per request.
The Space Bunny Alpha connection
About four days before the launch, stealth/space-bunny-alpha appeared on OpenRouter priced at $0 in and out for a preview window. This is a familiar pattern: a lab runs a free anonymous model to collect real-world usage and benchmark data without putting its name on something that might embarrass it, then ships a similarly specced model under its own brand days later. Z.ai did the same with “Ox Alpha,” revealed in September 2026 to be GLM-5.3 Flash — see Ox Alpha revealed as GLM-5.3 Flash.
Developers fingerprinted Space Bunny Alpha with the usual toolkit — tokenizer comparisons and error-behavior tests. One independent test sent 24 token-count probes through OpenCode’s Space Bunny route and found all 24 matched MiniMax models on the same gateway; a second measurement set matched 50 of 50. Shared traits: 1M context, forced thinking and comparable effort levels. MiniMax has not confirmed the link, and probably will not.
Where it fits in MiniMax’s lineup
| Model | Role | Price (per MTok) | Public benchmarks |
|---|---|---|---|
| M2.7 | Prior generation | ~10× cheaper than Claude Sonnet | Yes |
| M3 | Coding-only flagship (mid-2026) | $0.30 in / $1.20 out / $0.06 cached | 80.5% SWE-bench Verified, 59.0% SWE-Bench Pro |
| M3.1-Flash-Preview | Fast everyday coding, Sep 27, 2026 | Not published | None |
M3 is the reference. VentureBeat reported at its launch that M3 beat GPT-5.5 and Gemini 3.1 Pro on SWE-Bench Pro at roughly 5-10% of the cost. M3.1-Flash-Preview is MiniMax’s third distinct coding model release in under a year, each landing while outside labs were still benchmarking the last one. The strategy is explicit: undercut Anthropic and OpenAI by an order of magnitude on input tokens and ship fast enough that switching costs never accumulate. For M3 against the frontier, see MiniMax M3 vs Claude Fable 5 vs DeepSeek V4 Pro.
Honesty data point
In a September 27, 2026 web-extraction benchmark with a “Do not guess” instruction, MiniMax M3 invented 8 of 36 missing fields — mid-table, level with Claude Haiku 4.5 — for a $0.027 run. M3.1-Flash was not tested. If you use either for structured extraction, keep a verifier on; method in how to stop LLMs inventing fields in extraction.
Should you use it?
- Yes, to try: it is free inside MiniMax Code today, and the five-tier effort dial is worth understanding before it reaches the API.
- Not yet, for production: no model card means no knowledge cutoff, no safety notes, no rate limits, no SLA and no price to budget against. “Preview” in the name is doing real work.
- Watch for: an API listing with pricing (expect at or below M3’s $0.30/$1.20), a model card, and SWE-bench numbers. The day those appear, compare it against DeepSeek V4.1 Flash, MiMo-V2.6-Flash and Naive-N0.5-Flash in the open flash model comparison.
Last verified: September 28, 2026. M3.1-Flash-Preview has no official specification; details above come from MiniMax Code’s UI and MiniMax’s X post.