MiniMax has officially launched M3.1 Flash, the latest version of its “Flash” line of AI models built for speed and affordability rather than raw scale. The release matters because it promises near-flagship performance on agentic tasks, coding, computer use, and cybersecurity, while running nearly three times faster and costing 70% less than its predecessor to operate—a combination that could make advanced AI agents far more practical to deploy at scale.

From a secret codename to a public release

Before its official unveiling, M3.1 Flash existed in the wild under the anonymous codename “Space Bunny.” According to MiniMax, this early version quietly processed more than 200 trillion tokens while operating without a public identity, effectively serving as a large-scale real-world test before the formal launch. That kind of token volume suggests the model was already being used heavily in production-like conditions, giving MiniMax a substantial amount of usage data to refine the model ahead of its public debut.

What makes M3.1 Flash different

MiniMax describes M3.1 Flash as a “next-generation Flash model,” positioning it as the efficiency-focused counterpart to a more powerful “Pro” model that the company says will arrive later. The core differentiator is a deliberate trade-off favoring speed and cost over sheer model size, achieved through what MiniMax calls “more aggressive sparsity” in its inference process.

Sparsity, in simplified terms, refers to techniques that selectively activate only the parts of a neural network needed for a given task, rather than running the entire model for every request. This is similar in spirit to mixture-of-experts architectures used across the industry, though MiniMax does not specify the exact technical approach. The practical result, per the company, is output speeds of over 200 tokens per second (TPS) in everyday use, with median (P50) speeds consistently above 170 TPS—a significant jump for a model also claiming improved task performance.

Benchmark performance and pricing

MiniMax says M3.1 Flash reaches the level of leading models across several specific benchmarks:

• Automation-Bench, which tests complex agentic task execution
• DeepSWE, focused on real-world coding ability
• Agents’ Last Exam (ALE), which evaluates computer-use capability in tasks tied to real economic value

The company frames the headline result as performance “on par with top models at a very low task cost,” which it positions as a new standard for the performance-to-cost ratio in this category of model.

On pricing, MiniMax has published flat per-token rates: $0.15 per million input tokens, $0.60 per million output tokens, and a discounted $0.015 per million for cached input tokens. For agentic workloads specifically—where a model may need to take many sequential actions to complete a task—MiniMax states that overall usage cost is 70% lower than with its predecessor. For developers running agents that make repeated API calls, cached-input pricing in particular could meaningfully reduce costs when prompts share large amounts of repeated context, such as long system instructions or tool definitions.

Why speed and cost matter for agentic AI

Agentic AI systems—models that autonomously plan and execute multi-step tasks like browsing the web, writing and testing code, or controlling a computer interface—tend to be expensive and slow in practice because they often require many back-and-forth model calls to complete a single task. A model that is three times faster and 70% cheaper doesn’t just save money on a single request; it compounds across the dozens or hundreds of inference calls an autonomous agent might make while completing a workflow. This is likely why MiniMax emphasizes agentic benchmarks and computer-use capability alongside traditional coding tests, rather than focusing solely on general knowledge or reasoning benchmarks.

Cybersecurity is also listed among the model’s improved capabilities, though the source does not provide specific benchmark details for this category, so it’s worth treating that claim as a general improvement rather than a specifically quantified one.

How to access M3.1 Flash

MiniMax has made M3.1 Flash available immediately through three channels: the MiniMax API for developers building custom applications, the M Plan subscription service, and MiniMax Code, a dedicated coding-focused product. A desktop version of MiniMax Code is available for download directly from the company. Notably, MiniMax says model weights will be released at a later date, meaning self-hosting or fine-tuning the open weights is not yet an option at launch—access for now is limited to MiniMax’s own hosted infrastructure.

The company also confirms that a more capability-focused “Pro” model is still forthcoming, positioning Flash as the efficiency-oriented half of a two-tier strategy: Flash for speed and cost-sensitive use cases, and Pro for pushing raw intelligence further, likely at a higher cost and slower speed in exchange for stronger reasoning capabilities.

Frequently asked questions

Is M3.1 Flash open source?
Not yet. MiniMax states that model weights will be released later, so at launch the model is only accessible through MiniMax’s hosted API, M Plan, and MiniMax Code rather than for self-hosting.

What was “Space Bunny”?
Space Bunny was the anonymous codename for an early version of M3.1 Flash that MiniMax says processed over 200 trillion tokens before the model’s official public launch.

How is M3.1 Flash priced?
MiniMax lists flat rates of $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens, with overall agentic task costs cited as 70% lower than the previous model.

Source: MiniMax M3.1 Flash