MiniMax M3.1 Flash: Space Bunny Stealth Model Explained
MiniMax M3.1 Flash: Space Bunny Stealth Model Explained
For weeks, an anonymous model called Space Bunny quietly dominated AI leaderboards. On October 19, MiniMax admitted the bunny was theirs — and launched M3.1 Flash, a 428B agentic model priced at $0.15 per million tokens.
For the last month, AI Twitter had a running mystery: an anonymous model called Space Bunny was quietly eating every public leaderboard. It appeared out of nowhere, topped OpenRouter's volume charts for ten straight days, and disappeared — without anyone knowing who built it. On October 19, MiniMax ended the guessing game. Space Bunny was theirs: an early build of MiniMax M3.1 Flash, launched the same day.
And honestly, the launch was almost the least interesting part. What MiniMax actually pulled off was the debut of a new launch playbook: ship the model in secret, let 200 trillion tokens of real-world traffic validate it, and only then put your name on it. Let's break down what M3.1 Flash is, what it's actually good at, and why the stealth part matters more than the spec sheet.
The reveal: Space Bunny was MiniMax all along
MiniMax's launch post confirmed that Space Bunny was an earlier build of M3.1 Flash, saying the anonymous model had "topped global leaderboards with 200T+ tokens" before the company ever attached its name to it. The numbers from OpenRouter, which hosted the stealth run, are the concrete part: Space Bunny was its #1 model for 10 straight days and processed 56.7 trillion tokens — the most OpenRouter has ever recorded for a stealth model. It hopped off the platform on October 5, then returned on October 19 as M3.1 Flash, which OpenRouter is offering free for 72 hours.
The stealth run wasn't just a marketing stunt. OpenRouter said feedback from users during the anonymous period directly helped MiniMax ship a better model: time to first token improved 38% (from 1.95 seconds to 1.22 seconds), throughput rose 45% (from 70 to 100 tokens per second), and the error rate fell from 1.4% to 0.7%. That's the stealth playbook working as intended — free, unfiltered, real-world testing before the brand takes the risk.
What M3.1 Flash actually is
Under the hood, M3.1 Flash is a multimodal mixture-of-experts model: 428B total parameters with roughly 23B active per token, using sparse attention and a native visual encoder. It accepts text, image, and video input, produces text output, and carries a 1,000,000-token context window. MiniMax positions it for agentic work, coding, computer use, and cybersecurity workflows — squarely the "flash tier" of fast, cheap, capable models.
One design choice that will affect every developer using it: reasoning is always on and cannot be disabled. You get five effort levels — Low, Medium, High, Xhigh, and Max — with Max as the default. That sets a floor on both cost and latency for simple calls, since thinking tokens come out of your generated-token budget. On speed, MiniMax claims nearly 3x faster online inference than M3: median everyday output above 170 tokens per second (up to 200+), with the product page listing ~150 tok/s. Independent evaluator Artificial Analysis separately measured 120.1 tokens per second on MiniMax's API — fast, though short of the headline numbers.
The benchmarks: office-work star, hard-coding middleweight
Here's where you need to read the fine print. Every benchmark below is MiniMax-reported — no independent replication exists yet. With that caveat, the picture is interesting:
The split is clear: M3.1 Flash leads on structured agentic office work — Automation-Bench (multi-step workflows across SaaS tools) and BankerToolBench (multi-step planning, financial compliance, chained tool calls) — but it's mid-pack on hard coding. DeepSWE v1.1 at 73.7% is respectable yet trails DeepSeek's V4.1 Flash-1017 (77.0%) in MiniMax's own table, and Terminal Bench 4.0 at 38.1 vs. Opus 5's 51.8 is the score MiniMax would probably rather you skip. Also worth knowing: MiniMax says it ranks first among Chinese models on OSWorld V2 (61.7, computer use) and Agents' Last Exam (29.6).
All vendor scores above are MiniMax-reported; no independent replication exists yet, and benchmark numbers don't transfer cleanly across sources. What is independent: Artificial Analysis tested M3.1 Flash at max effort and recorded an Intelligence Index of 43 (up from M3's 29), 48% on Humanity's Last Exam (up from 39%), and one genuinely fun finding — the model answered 56% of factual-knowledge questions correctly, but 93% of its misses were flat-out wrong answers rather than "I don't know." This model would rather be confidently wrong than unsure. Keep that in mind before wiring it into anything that needs calibrated uncertainty.
The price is the real story
Here's the part that actually changes what developers can afford to build. M3.1 Flash is priced at $0.15 per 1M input tokens, $0.60 per 1M output tokens, and $0.015 per 1M cached-input tokens — a 90% discount on cached input, with no long-context surcharge. Per token, that's half of M3's standard rate, and MiniMax claims roughly 70% lower usage cost on agentic tasks at ~3x the speed. Artificial Analysis put the weighted cost per task at about $0.22, versus $0.51 for M3.
The cached-input pricing is the killer feature for agent builders: an agent that keeps re-reading the same codebase, document set, or conversation history across thousands of calls pays almost nothing for that repeated context. Combined with a 1M-token context window, this is priced for long-running agents, not one-shot chat.
And MiniMax isn't racing alone. DeepSeek's V4.1 Flash-1017 (launched days earlier) matches the $0.15/$0.60 rates off-peak with even cheaper cached input — though its peak-hour rates double — while GLM 5.3 and Kimi K3 crowd the same flash tier. Several Chinese labs have now shipped comparable agent-focused models within weeks of each other. When everyone's selling 1M context and tool use at commodity prices, the differentiator becomes reliability and latency under real load — which, funnily enough, is exactly what MiniMax's 200T-token stealth run was designed to prove.
What you can actually do with it this week
M3.1 Flash is live right now through the MiniMax API (model ID MiniMax-M3.1-Flash), the M Plan subscription, and the MiniMax Code desktop app. Two free windows are open: OpenRouter's 72-hour free period (started October 19) and Vercel's AI Gateway free access until October 22 (when MiniMax serves the request) — both revert to standard rates after.
The bigger date is October 23, when MiniMax releases the model weights. The license hasn't been disclosed yet, but the weights release is what lets independent researchers, smaller labs, and self-hosters verify MiniMax's claims directly and build fine-tuned variants — expect a wave of benchmark re-runs that may or may not match the vendor numbers.
If you're migrating from M3, watch out for two breaking changes: sending thinking.type=disabled or effort=none now returns an error (reasoning can't be turned off), and reasoning arrives in a separate reasoning_content field that may be absent — integrations that parsed inline <think> tags from M3 need updating. MiniMax also teased a Pro model "pushing the boundaries of intelligence" coming later, positioning Flash as the efficiency play in an ongoing lineup.
The bigger picture: stealth launches are the new launch strategy
Stealth-testing on public leaderboards isn't new — labs have shipped unlabeled checkpoints for ages to gather real usage data without brand risk. What MiniMax did differently was the reveal: instead of letting the connection stay as community speculation, it built Space Bunny's anonymous track record directly into the launch narrative. 200 trillion tokens of proof, name attached second.
Expect at least one more lab to confirm a previously anonymous leaderboard entry as its own flagship within the next quarter. The playbook works: bypass the skepticism that greets every new entrant, stress-test against live traffic, then convert months of anonymous dominance into a day-one launch story. The agent sandbox layer has become a utility — labs now compete on demonstrated throughput and reliability, not just benchmark charts.
Frequently asked questions
What was Space Bunny, exactly?
Space Bunny was an anonymous model that appeared on OpenRouter and other inference marketplaces in late September 2026, topping leaderboards without any lab claiming it. On October 19, MiniMax confirmed it was an early build of M3.1 Flash that had processed 200T+ tokens in stealth. On OpenRouter specifically, it was the #1 model for 10 straight days with 56.7 trillion tokens — the most OpenRouter has ever recorded for a stealth model.
How much does MiniMax M3.1 Flash cost — and can I try it free?
Flat API pricing is $0.15 per 1M input tokens, $0.60 per 1M output tokens, and $0.015 per 1M cached input tokens — half of M3's per-token rate, with no long-context surcharge. For free access: OpenRouter offered 72 hours free starting October 19, and Vercel's AI Gateway is free until October 22 (when MiniMax serves the request). Both revert to standard rates afterward.
Is M3.1 Flash open source?
Not yet — but it's close. MiniMax says the model weights arrive on October 23, 2026. The license hasn't been disclosed, so we don't yet know how permissive self-hosting and fine-tuning terms will be. Until then, it's API-only through MiniMax's own platform, the M Plan subscription, and MiniMax Code.
How good is it at coding, honestly?
Good at structured agentic work, middling at hard coding — based on vendor-reported scores. M3.1 Flash leads MiniMax's own Automation-Bench table (62.2) and ties Opus 5 on BankerToolBench (76.2), but scores 73.7% on DeepSWE v1.1 (trailing DeepSeek V4.1 Flash-1017's 77.0%) and just 38.1 on Terminal Bench 4.0 versus 51.8 for Opus 5. Independent testing by Artificial Analysis confirms solid reasoning gains over M3 but flags a tendency to give confidently wrong answers rather than admit uncertainty.
What breaks when I switch from M3 to M3.1 Flash?
Two things. First, reasoning is now always on: sending thinking.type=disabled or effort=none returns an error, and the effort field defaults to Max. Second, reasoning output arrives in a separate reasoning_content field (which may be absent) — if your M3 integration parsed inline <think> tags, it needs updating. Also note request ceilings and allowance accounting changed for subscription plans.
The bottom line
MiniMax didn't just launch a model — it proved you can debut at scale before you debut at all. 200 trillion tokens of anonymous traffic is a better validation set than any benchmark chart, and the stealth run visibly improved the product before launch day. For developers, the takeaway is simpler: a 428B-class agentic model at $0.15 per million tokens, free to try this week, with weights landing October 23.
The flash tier went from "budget option" to the most interesting place in AI this month — and Space Bunny is the reason everyone's paying attention.
Sources and further reading
- Unite.AI — MiniMax Launches M3.1 Flash, Revealing Space Bunny Stealth Model
- Tech Insider — MiniMax Space Bunny Reveal: M3.1 Flash, 200T Tokens
- Winbuzzer — MiniMax Releases M3.1 Flash AI Model With Substantial Cost Cuts and Better Performance
- MarkTechPost — MiniMax Just Launched M3.1 Flash: 428B MoE Agentic Model at $0.15/M Input
