MiniMax M3.1

As a next-generation Flash model, MiniMax M3.1 Flash delivers a generational leap in agentic tasks, coding, Computer Use and cybersecurity, while making online inference nearly three times faster and reducing usage costs by 70%. Space Bunny 🐰 on OpenRouter was an early version of MiniMax M3.1 Flash.

Model Summary

• Model Size: Built on a mixture-of-experts (MoE) architecture with 428 Billion total parameters and roughly 23 Billion active parameters.
• Context Window: Supports an expansive 1 Million token context window.
• Multimodal Inputs: Natively processes text, image, and video inputs, while outputting text.
• Tunable Thinking Depth: Developers can adjust reasoning settings dynamically from low to max effort mid-session.
• Speed: Reaches over 200+ tokens per second (TPS) in optimal everyday settings, with a consistent median (P50) speed of over 170+ TPS.

Benchmarks

Notes:

• AutomationBench v1.0.6: Evaluated on the 600 publicly released tasks (GitHub - zapier/AutomationBench: A benchmark for evaluating AI agents on realistic business workflows).
• BankerToolBench: Evaluated using the OpenCode 1.17.18 harness, with MiniMax-M2.7 as the judge model.
• GDP.pdf AA: Results are taken directly from Artificial Analysis (https://artificialanalysis.ai/).
• Terminal-Bench 4.0: using the Codex 0.149.1 harness.
• LOCA-bench 256K: Evaluated using the official implementation (GitHub - hkust-nlp/LOCA-bench: Benchmarking Language Agents Under Controllable and Extreme Context Growth).
• Agents’ Last Exam: Uses the official Overall setting. Since scores obtained on different evaluation subsets or through different evaluation methods are not directly comparable, this table includes only full-set results published on the official ALE leaderboard (https://agents-last-exam.org/leaderboard). Scores are left blank for models that have not undergone official evaluation or have reported only CLI evaluation results. Blank entries do not indicate a score of zero.
• OSWorld 2.0: Evaluated on the officially recommended 0808 Full release. Each model’s score is the highest reported at its release, except for DS-V4.1, whose score comes from an in-house reproduction.
• PostTrain Bench v1.1: Evaluated using the Claude Code 2.1.251 harness with a maximum context window of 512K tokens. All other settings follow the official evaluation setup (https://posttrainbench.com/).
• DeepSWE v1.1: Evaluated using the mini-SWE-agent 2.4.5 harness with a maximum context window of 512K tokens.
• NL2Repo-Bench: Evaluated using the Claude Code 2.1.233 harness with a maximum context window of 512K tokens. Evaluation cases involving reward hacking were assigned reward = 0. We also corrected two tasks with inconsistencies between their specifications and test cases. Reported scores are calculated on the corrected evaluation set.
• SWE Refactor Bench: Evaluated using the Claude Code 2.1.220 harness with a maximum context window of 200K tokens. All other settings follow the official evaluation setup (GitHub - Einsia/SWE-Refactor-Bench).
• CyberGym: Evaluated using the Claude Code 2.1.263 harness with a maximum context window of 1M tokens.
• BabyVision (w/ tools): Evaluated using Claude Code 2.1.285, with the minimum image side length set to 1,500 pixels.
• HLE (w/o tools): Evaluation is restricted to the 2,158 questions with text-only prompts. Reported scores are calculated on this subset.
• CharXiv RQ: Uses GPT-5.5 as the judge model.

Stronger Capabilities

M3.1 Flash improves across a broad range of benchmarks, reaching the level of leading models in complex agentic tasks (cowork), computer use, software development (coding), and cybersecurity.

Complex Tasks and Automation

• M3.1 Flash ranks among the top performers on AutomationBench. It can complete real enterprise workflows across SaaS systems, solving practical problems in sales, marketing, operations, customer service, finance, and HR from end to end.
• Scoring 76.2 on BankerToolBench, it excels at complex real-world challenges such as multi-step planning, maintaining consistency across documents and files, financial compliance, and chained tool calls.

Coding and Software Engineering

• It achieves 73.7% on DeepSWE v1.1, placing it alongside top models. DeepSWE focuses on complex software tasks that require models to independently understand code, develop features, fix issues, and validate their work through testing. M3.1 Flash demonstrates strong execution across long sequences of steps, supporting complex software development workflows at a lower cost.

Computer Use

• It scores 61.7 points on OSWorld 2.0, 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. It ranks first among Chinese models on both benchmarks.
• M3.1 Flash demonstrates 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.

Cybersecurity

• It achieves a strong score of 82% on CyberGym, demonstrating leading capabilities in vulnerability discovery and validation.
• In in-depth evaluations by our internal security experts, the model performed strongly in code auditing, vulnerability patching, penetration testing, malware analysis, and reverse engineering, enabling efficient, low-cost support for building complex security systems.

Agentic Tasks

Across three benchmarks — AutomationBench for complex agentic task execution, DeepSWE for real-world coding, and Agents’ Last Exam (ALE) for computer use in real-world tasks with economic value — 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.