Recommended Articles
Tech & AI

MiniMax M3.1 Flash Launches, Ending the "Space Bunny" Stealth Run

By Angeled · October 20, 2026 · Tech & AI
MiniMax M3.1 Flash Launches, Ending the "Space Bunny" Stealth Run
MiniMax M3.1 Flash Launches, Ending the "Space Bunny" Stealth Run

On October 19, 2026, Chinese AI lab MiniMax officially launched MiniMax M3.1 Flash, its next-generation fast multimodal model — and confirmed one of the industry's worst-kept secrets: the viral anonymous model "Space Bunny" that topped public leaderboards was an early build of the new release. The announcement, covered by Unite.AI, Forkast, and others, marks one of the most deliberate "stealth-to-scale" model launches of the year.

The Space Bunny stealth run

For roughly four weeks, an anonymous model operating as "Space Bunny" on public routing platforms such as OpenRouter quietly dominated developer attention. OpenRouter, which hosted the stealth run, said Space Bunny was its number-one model for 10 consecutive days and processed 56.7 trillion tokens on its platform alone — the most of any stealth model it has run. MiniMax said the model processed more than 200 trillion tokens in total across public leaderboards under the alias.

MiniMax confirmed the connection in a short video posted to its official accounts featuring a carrot-shaped rocket lifting off alongside the caption "Welcome back, Space Bunny." The company had first shipped an M3.1 Flash Preview in September; the October 19 release is the general-availability version.

What the launch data shows

OpenRouter published the stealth-run telemetry that MiniMax collected during the anonymous period, framing it as real-world validation rather than lab benchmarks:

  • Time to first token improved 38%, from 1.95 seconds to 1.22 seconds
  • Throughput rose 45%, from 70 to 100 tokens per second
  • Error rate fell from 1.4% to 0.7%

MiniMax's own self-reported benchmark scores include 73.7% on DeepSWE v1.1, 61.7 points on OSWorld V2, 29.6 points on ALE, 76.2 on BankerToolBench, and 82% on CyberGym — results the company describes as on par with top-tier models for agent tasks. The company also claims M3.1 Flash ranks first among Chinese models on OSWorld V2 and ALE.

Pricing and availability

M3.1 Flash is available now through the MiniMax API, the company's M Plan subscription, and MiniMax Code (a desktop app). MiniMax describes the model as a fast multimodal frontier model built for complex multi-step agentic work, coding, and cybersecurity, with a 1-million-token context window and up to 524,000 tokens of max output.

Pricing is flat per million tokens: $0.15 input, $0.60 output, and $0.015 cached input — a 90% discount on repeated context, which is aimed directly at agent workloads that re-read large codebases or document sets. MiniMax claims a 70% reduction in per-task cost versus its predecessor. OpenRouter also listed the model free for a 72-hour launch window starting October 19.

Model weights are promised for a later release, and a "Pro" model that MiniMax says pushes the boundaries of intelligence will arrive later — no dates were given for either.

Market reaction

MiniMax shares, listed on the Hong Kong Stock Exchange (0100.HK), gained 1.45% to close at HK$182.1 on the news. The launch lands alongside competing fast lightweight models from DeepSeek (V4.1 Flash) and Z.ai (GLM 5.3 Flash), forming what observers describe as a fast-moving Chinese trio in the high-speed, lightweight model tier.

Why it matters for founders

M3.1 Flash is a price signal as much as a product launch: sub-$0.20-per-million input pricing paired with 1M context means multi-agent pipelines that were cost-prohibitive in 2025 are now deployable at scale. Founders building agent products should benchmark their inference spend against this tier — if your agent stack costs 3x what M3.1 Flash prices suggest, your unit economics may already be obsolete.

Why it matters for investors

The stealth-launch playbook — anonymous leaderboard runs that generate months of organic developer demand before branding — is now proven infrastructure for go-to-market in AI. Investors evaluating model startups should look for this pattern: 200 trillion tokens of real usage is far more convincing traction than any launch-day demo. It also compresses pricing power across the lightweight-model tier, which pressures margins for every agent-platform company priced above it.

Angeled.com connects founders with investors.

Sources: Unite.AI, Forkast, Glonce, aifree.dev, tech-insider.org, cellcog.ai

Read Next