MiniMax M3.1 Flash

MiniMax M3.1 Flash is a next-generation Flash model from MiniMax, officially rolled out as a model built for agentic tasks, coding, Computer Use and cybersecurity. It is aimed at developers, teams and enterprises that need to run agents and software workflows at scale, whether through the MiniMax API, the M Plan subscription, or MiniMax Code, the company's desktop agent product. Its stated purpose is to deliver a generational leap in agentic capability while making online inference nearly three times faster and reducing usage costs by 70% compared with its predecessor, so that complex, multi-step work becomes affordable to run rather than a premium experiment. The model has a public backstory that MiniMax uses to frame its development: Space Bunny, an early version of MiniMax M3.1 Flash, processed more than 200 trillion tokens while operating anonymously before the model was formally released. MiniMax positions Flash as an ongoing exploration of the balance between efficiency and capability, distinct from Pro, which the company describes as pushing the boundaries of intelligence and which it says will arrive later. The problem M3.1 Flash addresses is the cost and latency of running agents that must take many steps, call tools repeatedly, and carry work across documents, files, and applications. Those workflows multiply token consumption, so speed per token and price per token directly determine whether an agent is practical to operate at all. For complex tasks and automation, M3.1 Flash ranks among the top performers on AutomationBench-AA and is described as able to complete real enterprise workflows across SaaS systems, solving practical problems in sales, marketing, operations, customer service, finance, and HR from end to end. It also scores 76.2 on BankerToolBench, where MiniMax says it excels at complex real-world challenges such as multi-step planning, maintaining consistency across documents and files, financial compliance, and chained tool calls. These capabilities matter because enterprise automation rarely fails on a single answer; it fails when an agent loses track of state between steps, misaligns two documents, or drops a tool call mid-sequence. In coding and software engineering, the model achieves 73.7% on DeepSWE v1.1, placing it alongside top models on a benchmark that focuses on complex software tasks requiring a model to independently understand code, develop features, fix issues, and validate its work through testing. MiniMax reports strong execution across long sequences of steps, supporting complex software development workflows at a lower cost. That combination is significant for teams running coding agents continuously, since the value of an agent that can plan, edit, test, and re-check its own work depends heavily on how many steps it can sustain before cost or latency becomes prohibitive. For Computer Use, M3.1 Flash scores 61.7 points on OSWorld V2, which measures AI computer operation and cross-application task execution, and 29.6 points on Agents' Last Exam (ALE), which evaluates performance on real professional work and complex tasks; MiniMax states it ranks first among Chinese models on both benchmarks. The company describes the model as demonstrating strong capabilities in interacting with real environments, using tools, and executing long sequences of tasks, moving beyond understanding a task to taking action and delivering results. In cybersecurity, it achieves 82% on CyberGym for vulnerability discovery and validation, and in in-depth evaluations by MiniMax's internal security experts it performed strongly in code auditing, vulnerability patching, penetration testing, malware analysis, and reverse engineering, enabling efficient, low-cost support for building complex security systems. Overall, MiniMax says it further optimized M3.1 Flash's inference efficiency through more aggressive sparsity. In everyday use, output speeds can reach 200+ tokens per second, with median (P50) speeds consistently reaching 170+ tokens per second. Across three benchmarks, Automation-Bench for complex agentic task execution, DeepSWE for real-world coding, and Agents' Last Exam for computer use in real-world tasks with economic value, MiniMax reports that M3.1 Flash delivers performance on par with top models at a very low task cost, setting a highly competitive new standard for performance and cost. The underlying methodology is stated plainly: aggressive sparsity for faster online inference, paired with a deliberate trade-off between efficiency and raw intelligence. The benefits flow directly from those numbers. For agentic tasks, the model's overall usage cost is 70% lower than its predecessor, and pricing is flat per 1M tokens at $0.15 input, $0.60 output, and $0.015 cached input. Faster output speeds mean agents return results in less wall-clock time, while cached input pricing lowers the cost of repeatedly sending the same context, which is common in long-running agent sessions. Together, higher throughput and lower per-token cost change the economics of running agents continuously rather than occasionally, and they let teams scale from a single workflow to many without multiplying spend proportionally. Concrete use cases described by MiniMax include end-to-end enterprise workflows across SaaS systems in sales, marketing, operations, customer service, finance, and HR; financial work involving multi-step planning, document and file consistency, financial compliance, and chained tool calls; software development in which the model independently understands code, develops features, fixes issues, and validates its work through testing; computer-use scenarios requiring cross-application execution on real environments; and security work spanning vulnerability discovery, code auditing, vulnerability patching, penetration testing, malware analysis, and reverse engineering. M3.1 Flash is available through the MiniMax API, M Plan, and MiniMax Code, with MiniMax Code available as a desktop download. MiniMax states that model weights will be released later, and that a Pro model will also arrive later. Pricing is published as a single flat rate per 1M tokens rather than tiered by context length: $0.15 for input, $0.60 for output, and $0.015 for cached input. In summary, MiniMax M3.1 Flash is positioned as a fast, low-cost model for the work agents actually do: multi-step automation, real coding tasks, computer operation, and security analysis. Its value proposition rests on measured claims, roughly three times faster online inference, 200+ tokens per second output, and 70% lower usage cost than its predecessor, delivered alongside benchmark results MiniMax says place it near top models in agentic execution, coding, and computer use.