MiniMax M3.1 Flash

Architectural Overview & Capability Profile

MiniMax M3.1 Flash (commonly known as minimax-m31f) is MiniMax’s flagship open-weights Mixture-of-Experts (MoE) foundation model. Engineered specifically to tackle the dual bottlenecks of high inference cost and latency in long-running agent loops, M3.1 Flash deploys 428 billion total parameters with 23 billion active parameters per token. Across RankLLMs’s calibrated evaluation framework, MiniMax M3.1 Flash secures an overall rating of 58.6 / 100, establishing itself as one of the fastest and most cost-effective agentic models worldwide.

Distributed under an open-weights architecture with permissive self-hosting and inference support across vLLM, Ollama, and Hugging Face, M3.1 Flash couples an expansive 1,000,000-token input context window with an unprecedented 512,000-token maximum output budget. This allows the model to maintain deep chain-of-thought scratchpads and execute multi-thousand-line codebase refactors in a single pass without context truncation.

Benchmark Deep Dive: CoWork, Coding & Tool Automation

In verified cross-lab evaluations conducted against DeepSeek-V4.1-Flash, Kimi K3, and frontier proprietary models, MiniMax M3.1 Flash demonstrates standout proficiencies across five key domains:

1. Autonomous Enterprise CoWork: MiniMax M3.1 Flash captures the global lead on AutomationBench v1.0.6 at 62.8%, surpassing DeepSeek-V4.1-Flash-1017 (60.5%), Claude Opus 5 (50.3%), and GPT-5.6 Sol (45.8%). In specialized finance and enterprise workflows, it matches top closed-source frontier architectures with 76.2% on BankerToolBench and 60.3% on LOCA-Bench 256K.
2. Long-Horizon Coding & Repo Fixing: On the rigorous DeepSWE v1.1 benchmark evaluated using mini-SWE-agent, M3.1 Flash registers 73.7%, rivaling Claude Opus 5 (74.0%) and GPT-5.6 Sol (73.0%). In complex repository understanding tasks, it logs 61.2% on NL2Repo-Bench and 23.0% on SWE Refactor Bench.
3. Computer Use & Desktop Navigation: Evaluated on the official OSWorld 2.0 release, M3.1 Flash achieves 61.7%, demonstrating smooth coordination across terminal and GUI interfaces.
4. Cybersecurity & Safety: Scoring 82.2% on CyberGym via Claude Code 2.1 harnesses, M3.1 Flash safely identifies, isolates, and patches real-world security vulnerabilities.
5. Multimodal Reasoning: M3.1 Flash achieves 91.0% on BabyVision (with tools) and 82.5% on CharXiv RQ, processing dense visual schematics and research figures with high accuracy.

Inference Throughput & Cost-per-Task Economics

MiniMax M3.1 Flash operates at an extreme cost and latency advantage:

• Streaming Throughput: Yields 185 tokens per second at steady state, delivering a 3x throughput improvement over prior-generation MiniMax M3.
• Blended Pricing: At $0.15 / 1M input tokens and $0.60 / 1M output tokens (with $0.015 / 1M KV cache read), the blended 3:1 token expenditure comes to just $0.26 / 1M tokens. Running a 50-step autonomous agent harness costs under $0.05 per resolved issue, representing an 85% expenditure reduction relative to proprietary frontier options.