
MiniMax-M3.1-Flash
API Overview
MiniMax-M3.1-Flash (code-named “Space Bunny”) is the new-generation flagship MoE deep-thinking and long-context Agent model released by MiniMax. As a major leap forward from the M3 to the M3.1 generation, M3.1 Flash achieves agentic automation and code-generation capabilities approaching those of top-tier frontier models, all while maintaining the ultra-low API pricing characteristic of the Flash tier. The model natively integrates persistent adaptive deep thinking and multimodal visual perception, delivering outstanding performance on automation benchmarks such as Automation-Bench, DeepSWE code refactoring, and computer desktop/web operations. Coupled with its native 1M+ token context window and advanced prompt caching mechanism, it serves as the gold-standard industrial-grade foundation for enterprises building cost-effective AI agents, automating complex workflows, and analyzing large codebases.
───────────────────────────────────────────────────────────────────
Core Capabilities
State-of-the-Art Agentic Coding and Workflow Automation: Scores 62.2% on Automation-Bench and 73.7% on DeepSWE; excels at unattended long-form code debugging, multi-tool invocation, and automated system interaction.
Persistent Reasoning Chains and Adaptive Inference: Employs a five-level inference hierarchy, defaulting to maximum reasoning depth to ensure high rigor and self-checking/correction in complex, multi-step reasoning tasks.
Multimodal Visual Understanding and 1M+ Token Context: Natively supports an ultra-long context window of 1,000,000 tokens, directly accepting UI screenshots, charts, and engineering drawings to execute precise vision-driven automated decisions.
Ultra-Cost-Effectiveness and Deep Caching: Features an exceptionally low Cache Read rate—down to 10%—significantly reducing redundant context consumption in multi-turn agent loops, leading to a substantial reduction in overall task compute costs.
Playground
Log in to explore more features! Click to Log In