MiniMax M3.1 Flash Officially Revealed: 428B-Parameter 'Space Bunny' Agentic AI Cuts Costs 70%, Ignites APAC Price War
From 'Space Bunny' to M3.1 Flash: A Carefully Orchestrated Reveal
On October 19, 2026, Chinese AI unicorn MiniMax officially launched the M3.1 Flash model, simultaneously solving a mystery that had circulated in the AI community for weeks: the enigmatic model quietly testing on the OpenRouter platform under the codename "Space Bunny" was M3.1 Flash all along.
This weeks-long "stealth testing" was no accident. Before the official launch, M3.1 Flash processed over 200 trillion tokens anonymously on OpenRouter, accumulating vast real-user data while maintaining an air of mystery. This "test first, reveal later" strategy has pioneered a new launch paradigm in the AI industry.
Technical Specifications: An Efficiency Revolution for Agentic AI
Core Architecture
M3.1 Flash is an upgraded version of MiniMax's M3 model launched in June 2026, employing a Mixture-of-Experts (MoE) architecture:
| Specification | M3.1 Flash | M3 (Previous) |
|---|---|---|
| Total Parameters | 428 billion | Undisclosed |
| Inference Speed | >200 tokens/sec | Baseline |
| Speed Improvement | 3x faster | — |
| Context Window | 1 million tokens | 1 million tokens |
| Supported Modalities | Text, Image, Video | Text, Image |
Agentic Capability Benchmarks
M3.1 Flash demonstrates strong performance across multiple agentic AI benchmarks:
- OSWorld V2 (computer use): 61.7
- DeepSWE v1.1 (autonomous coding): 73.7%
- BankerToolBench (multi-step planning/financial compliance): 76.2
- CyberGym (vulnerability reproduction/security): 82%
On Artificial Analysis's Intelligence Index, M3.1 Flash scored 42, a significant improvement from M3's score of 29. While still trailing top-tier US frontier models, it offers a compelling cost-performance ratio.
70% Cost Reduction: Redefining the Economics of Agentic AI
Pricing Structure
MiniMax is disrupting the market with aggressive pricing:
| Billing Type | M3.1 Flash | Market Average |
|---|---|---|
| Input Tokens | $0.15/million | Competitors avg $0.5+ |
| Output Tokens | $0.60/million | Competitors avg $2.0+ |
| Cached Input | $0.015/million | — |
In the Chinese market, pricing is even more aggressive: 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.1 yuan per million for cached reads.
Per-Task Cost Comparison
MiniMax claims that compared to M3, M3.1 Flash reduces average agentic task costs from $0.51 to $0.22 per task — a 57% reduction (officially claimed as 70%, measured at 57%). This gap primarily comes from more aggressive sparse inference strategies that dramatically reduce computational resource consumption while maintaining output quality.
Strategic Significance for the Asia-Pacific Market
China's AI Price War Reaches New Heights
The launch of M3.1 Flash is the latest chapter in China's AI industry "price war." In the Asia-Pacific region, MiniMax faces fierce competition from domestic rivals like Zhipu AI, Alibaba Tongyi, and Xiaomi, as well as international competitors like Google Gemini Flash and Anthropic Haiku.
According to Caixin Global reporting, M3.1 Flash's pricing in the Chinese market is approaching the "cost floor," indicating MiniMax is pursuing a strategy of trading scale for market share.
Open-Source Plans and Ecosystem Building
MiniMax announced plans to release M3.1 Flash's open weights on October 23, 2026, a move that will further accelerate its ecosystem building in the Asia-Pacific region. The open-source strategy not only attracts developer communities but also enables enterprises to deploy models locally, satisfying data sovereignty requirements.
Compliance Considerations for Enterprise Deployment
For enterprises operating in the Asia-Pacific region, MiniMax offers direct commercial agreements supporting customized deployments and compliance configurations. This is particularly important for industries like finance and healthcare that need to comply with local data regulations.
Technical Highlight: The Breakthrough in Sparse Inference
The core of M3.1 Flash's performance improvement lies in its "more aggressive sparse inference" strategy. In the MoE architecture, each inference only activates a subset of expert networks, and M3.1 Flash further optimizes the expert selection algorithm to minimize computation while maintaining output quality.
The trade-off of this approach is a higher hallucination rate (93%), primarily because the model tends to attempt answering all questions rather than declining when uncertain. For enterprise applications requiring high accuracy, this characteristic warrants special attention, and it's recommended to use validation mechanisms alongside the model.
Multimodal Capabilities: Agentic AI Beyond Text
M3.1 Flash supports multimodal inputs including text, images, and video, giving it unique advantages in scenarios such as:
- Visual Agent Tasks: Analyzing screenshots, operating interfaces, understanding charts
- Video Content Processing: Automatic summarization, content moderation, scene analysis
- Document Intelligence: PDF parsing, table extraction, contract review
Availability and Deployment Options
As of October 2026, M3.1 Flash is available through:
- MiniMax API Platform: API access for developers
- M Plan Subscription: Subscription service for individual users
- MiniMax Code Desktop Application: Local tool for developers
- Open Weights (planned): Supporting local deployment
Conclusion: The Cost Revolution of the Agentic AI Era
The launch of MiniMax M3.1 Flash is not just a technical milestone but an important signal for the commercialization of agentic AI. When the per-task cost of agentic AI drops to $0.22, the economic viability of large-scale automated workflows improves dramatically.
For enterprises in the Asia-Pacific region, M3.1 Flash offers a choice that balances performance and cost. With the release of open weights, the possibility of local deployment will further lower the barrier to AI adoption, driving the popularization of agentic AI across the Asia-Pacific region.
