International access: https://4sapi.com
Domestic access: https://4sapi.org
Abstract
On October 19, MiniMax officially launched its next-generation Flash model, M3.1 Flash. This release marks the public reveal of the previously anonymous "Space Bunny" model, a system that processed more than 200 trillion tokens during its period of quiet production operation. Built around an optimized sparse attention architecture, M3.1 Flash targets high throughput and low operational cost for agent workloads. It delivers sustained token generation speed above 200 tokens per second (TPS), with a median stable throughput of 170+ TPS, bringing nearly a 3x uplift in online inference speed. The model natively supports a 1,000,000-token context window and multimodal inputs covering text, images and video, alongside five adjustable effort levels to tune reasoning depth. Benchmark results show competitive performance on agent automation, software engineering, computer-use and cybersecurity tasks, rivaling flagship closed models at a fraction of the cost. This article breaks down its technical architecture, pricing structure, benchmark outcomes, industry positioning alongside DeepSeek V4.1 Flash-1017, and practical deployment considerations for enterprise agent pipelines.
1. Introduction: The Unveiling of Space Bunny
For an extended period, the model known internally as Space Bunny ran anonymously in production, handling a massive token volume exceeding 200 trillion tokens of real-world user traffic. This large-scale online serving run acted as a live validation phase, gathering real distribution data and refining model behavior before public release. The official launch of M3.1 Flash formally identifies Space Bunny as the predecessor of this new Flash model.
Flash-class large language models were once positioned for lightweight chat and simple auxiliary tasks. The landscape is now shifting rapidly. New Flash variants are entering production-grade enterprise workflows, handling complex agent automation, end-to-end coding, computer interface control and security evaluation workloads that previously required expensive flagship models. MiniMax’s M3.1 Flash is one representative of this transition. Combined with DeepSeek V4.1 Flash-1017, the pair creates what industry observers term the domestic Flash "twin stars", compressing the cost of sophisticated single tasks down to just a few units of currency. This shift redefines the economics of building long-running agent systems.
This article dissects the core technical design of M3.1 Flash, its throughput and pricing parameters, benchmark scores across key agent evaluation suites, and what this release signals for the broader LLM serving ecosystem. It also addresses practical engineering considerations when integrating high-speed Flash models into multi-model agent stacks.
2. Architecture, Throughput and Pricing
M3.1 Flash’s core optimization relies on an aggressive sparse attention design. Sparse attention reduces the computational overhead of long-context inference by limiting attention calculation to relevant token segments, rather than computing full pairwise attention across the entire context window. This design choice is the foundation for its high generation throughput.
2.1 Throughput Metrics
In live production serving environments, M3.1 Flash can deliver peak output speeds exceeding 200 TPS. Median stable throughput maintains above 170 TPS under typical workloads. Compared with prior generations of MiniMax models, online inference speed increases by nearly three times. This level of throughput is critical for agent workflows. Multi-step agent loops involve repeated model calls, and high latency quickly compounds across dozens or hundreds of tool-calling iterations. Higher token throughput reduces wall-clock runtime for long agent trajectories and improves user experience in interactive automation scenarios.
2.2 Pricing and Token Cost
MiniMax sets a flat, all-day pricing scheme without peak-hour surcharges, a key advantage for continuous agent workloads that run at all hours. The published pricing is:
- Input tokens: $0.15 per million tokens
- Output tokens: $0.60 per million tokens
- Cached input tokens: $0.015 per million tokens
Converted to approximate local currency equivalents, this lands around 1 RMB per million input tokens and 4 RMB per million output tokens. For agent workloads, the comprehensive usage cost drops by 70% compared with earlier model generations. The massive cost reduction changes the feasibility of running high-volume agent automation at enterprise scale. Previously, complex continuous agent workflows incurred prohibitive token bills; M3.1 Flash brings these workloads within practical budget limits.
2.3 Native Capabilities and Reasoning Tuning
The model ships with native support for a 1,000,000-token context window. It accepts multimodal inputs: plain text, static images and video content. Multimodal support is essential for modern computer-use agents that must interpret screenshots, visual interfaces and video frames to operate graphical desktop environments.
Another notable feature is five configurable effort levels. These settings adjust the model’s reasoning depth. Developers can select lower effort for fast, simple tasks to cut latency and token consumption, or enable higher effort for complex planning, mathematical reasoning and multi-step agent decision making. This granular control gives operators a knob to balance speed, cost and task performance on a per-request basis.
3. Benchmark Performance Across Agent-Focused Evaluation Suites
The benchmark results demonstrate that M3.1 Flash is not merely a fast chat model. It achieves competitive scores on demanding agent, coding, computer-use and cybersecurity benchmarks, matching flagship closed models on many tasks.
| Benchmark | Score | Notes |
|---|---|---|
| Automation-Bench | 62.8 | Ranked first for complex agent workflows |
| DeepSWE v1.1 (max mode) | 73.7 | Matches flagship GPT-5.6 Sol |
| OSWorld 2.0 (Computer Use) | 61.7 | Near GPT-5.6 Sol performance |
| ALE | 29.6% pass rate | Lowest total cost (~739 CNY) to complete all 152 questions among over 20 evaluated models |
| CyberGym | 82.2 | Enters the top tier for network security evaluation |
Automation-Bench measures end-to-end complex agent workflow execution. M3.1 Flash’s 62.8 score takes first place on this benchmark, proving its strength for chained multi-step automation tasks. For software engineering tasks on DeepSWE v1.1, its max-effort score of 73.7 reaches parity with GPT-5.6 Sol, a top-tier flagship model.
On computer-use benchmarks, OSWorld 2.0 scores 61.7. The ALE benchmark, which tests agent ability to operate software interfaces, records a 29.6% pass rate. A standout economic result comes from the ALE test set: running all 152 evaluation questions costs approximately 739 RMB. This is the lowest total cost among more than twenty competing models. The combination of strong pass rate and minimal inference cost represents a compelling value proposition for teams building desktop automation agents.
In cybersecurity testing on CyberGym, the model scores 82.2, placing it in the industry’s top tier. This shows that the Flash architecture retains strong reasoning for specialized security tasks, not just general automation and coding.
Collectively, these benchmarks illustrate a key shift. Flash models are no longer limited to simple retrieval-augmented generation or basic chat. They can handle high-complexity agent tasks once reserved for large, expensive flagship models.
4. Industry Positioning: The Flash Twin Stars and The Rise of Flash as Primary Production Workhorses
The release of M3.1 Flash arrives shortly after DeepSeek V4.1 Flash-1017, launched two days prior. The pair has been referred to as the domestic Flash twin stars. Together, they push down the cost ceiling for enterprise workflows, programming and computer-use automation. Complex tasks that once required flagship model pricing now run at a per-task cost of just one or two RMB.
This milestone marks a broader transition in LLM deployment. Flash-class models are moving from auxiliary lightweight assistants to primary production workhorses. For many agent use cases, teams can now adopt Flash models as their default backend instead of reserving them only for trivial low-stakes tasks. Flagship models remain useful for the most complex reasoning edge cases, but Flash models absorb the bulk of routine agent execution volume.
Product Availability Roadmap
At launch, M3.1 Flash is accessible through API endpoints, the M Plan platform and MiniMax Code. Model weights are scheduled for release at a later date. A higher-end Pro variant of the model is also in active development.
The staged rollout gives API-first teams immediate access to evaluate and integrate the model, while teams requiring local deployment can wait for weight release. The upcoming Pro version indicates that MiniMax will continue expanding the Flash product family to serve higher-complexity workloads.
5. Operational Challenges for Multi-Flash Agent Deployments
While throughput and cost are attractive, engineering teams must address several practical challenges when building production systems around M3.1 Flash and similar high-speed Flash models.
First, high throughput does not eliminate token consumption. Even with lower per-million-token pricing, continuous long-running agent loops can accumulate large token volumes. The 70% cost reduction versus prior models is significant, but cost monitoring and request throttling remain necessary.
Second, multimodal input adds complexity. Image and video preprocessing, payload resizing and media encoding must be handled before sending requests to the model API. Teams need standardized pipelines to convert screenshots and video frames into compatible input formats.
Third, multi-model routing becomes essential when mixing multiple Flash models and flagship backends. Many agent platforms route simple tasks to Flash models and escalate hard reasoning tasks to larger models. An API gateway simplifies managing multiple model endpoints, authentication keys, rate limits and usage logging. 4sapi functions as an API gateway to centralize credential management, traffic routing and consumption tracking for multi-model agent stacks. For teams running both M3.1 Flash and DeepSeek V4.1 Flash-1017, unified routing reduces the overhead of separate API integrations for each model provider.
Fourth, benchmark performance does not guarantee identical behavior in custom production distributions. Benchmark suites are curated test sets. Real enterprise workflows contain unique domain rules, edge cases and unexpected inputs. Teams must implement internal evaluation pipelines to validate agent behavior before full production rollout.
6. Conclusion
MiniMax M3.1 Flash, the official public identity of the previously anonymous Space Bunny model, demonstrates how Flash-class models are reshaping the economics of AI agent development. Built with sparse attention, it delivers 170+ median TPS and nearly 3x faster online inference, while cutting agent workload costs by 70%. Its 1M-token native context window, multimodal text-image-video input, and five-level reasoning effort controls provide flexible knobs for agent builders.
Benchmark results are highly competitive. M3.1 Flash tops Automation-Bench, matches GPT-5.6 Sol on DeepSWE v1.1 coding tasks, delivers near-flagship performance on computer-use benchmarks and enters the top tier on CyberGym security evaluation. The ALE benchmark highlights its exceptional cost efficiency, with the lowest total cost across more than twenty competing models.
Paired with DeepSeek V4.1 Flash-1017, M3.1 Flash forms the domestic Flash twin stars. This release signals that Flash models are evolving from lightweight side models into primary production backends for enterprise agent automation, coding and computer-use tasks. The model is live on APIs at launch, with weight release and a Pro version planned for later.
For engineering teams building multi-model agent systems, centralized API routing and consumption tracking streamline the process of evaluating and deploying high-performance Flash models like M3.1 Flash into live production pipelines.
