The evaluation results below are taken from the model card of the original model “XiaomiMiMo\/MiMo-V2.6-Flash-RL”. Since this model is a GGUF quantized version, slight accuracy variations may occur due to quantization. Note that due to the high number of columns in the table, it has been narrowed down to focus on the original model and major comparative models (numerical values are kept as in the original text).<\/p>\n
\n
\u2192 Scroll horizontally to see all columns<\/p>\n
\n\n\n| Benchmark<\/th>\n | MiMo-V2.6 Flash<\/th>\n | MiMo-V2.6 Pro<\/th>\n | MiMo-V2.5 Pro<\/th>\n | Claude Opus 5<\/th>\n | GPT-5.6 Sol<\/th>\n<\/tr>\n<\/thead>\n |
\n\n| Code Agent<\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n<\/tr>\n |
\n| DeepSWE v1.1<\/td>\n | 67.9<\/td>\n | 71.9<\/td>\n | 19.0<\/td>\n | 74.0<\/td>\n | 73.0<\/td>\n<\/tr>\n |
\n| ProgramBench<\/td>\n | 26.0<\/td>\n | 26.5<\/td>\n | 12.5<\/td>\n | 37.0<\/td>\n | 25.0<\/td>\n<\/tr>\n |
\n| MiMo Code Bench<\/td>\n | 61.2<\/td>\n | 63.2<\/td>\n | 40.4<\/td>\n | 68.6<\/td>\n | 59.3<\/td>\n<\/tr>\n |
\n| General Agent<\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n<\/tr>\n |
\n| AutomationBench v1.0.6<\/td>\n | 52.3<\/td>\n | 53.1<\/td>\n | 16.0<\/td>\n | 50.3<\/td>\n | 45.8<\/td>\n<\/tr>\n |
\n| Toolathlon-Verified<\/td>\n | 73.6<\/td>\n | 76.9<\/td>\n | 49.1<\/td>\n | 80.6<\/td>\n | 74.9<\/td>\n<\/tr>\n |
\n| Agents\u2019 Last Exam<\/td>\n | 27.6<\/td>\n | 31.6<\/td>\n | 13.2<\/td>\n | 31.6<\/td>\n | 30.8<\/td>\n<\/tr>\n |
\n| Terminal Bench 4.0<\/td>\n | 28.8<\/td>\n | 34.9<\/td>\n | 1.5<\/td>\n | 49.0<\/td>\n | 39.9<\/td>\n<\/tr>\n |
\n| Terminal Bench 2.1<\/td>\n | 87.6<\/td>\n | 89.9<\/td>\n | 65.2<\/td>\n | 89.1<\/td>\n | 88.8<\/td>\n<\/tr>\n |
\n| OSWorld-Verified<\/td>\n | 80.8<\/td>\n | 82.0<\/td>\n | –<\/td>\n | 83.4<\/td>\n | 83.0<\/td>\n<\/tr>\n |
\n| JobBench<\/td>\n | 61.2<\/td>\n | 62.0<\/td>\n | 25.0<\/td>\n | 65.7<\/td>\n | 45.4<\/td>\n<\/tr>\n |
\n| Cybersecurity<\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n<\/tr>\n |
\n| CyberGym<\/td>\n | 95.1<\/td>\n | 94.0<\/td>\n | 40.0<\/td>\n | –<\/td>\n | –<\/td>\n<\/tr>\n |
\n| MiMo Cyber Bench<\/td>\n | 77.2<\/td>\n | 80.2<\/td>\n | 0.0<\/td>\n | –<\/td>\n | –<\/td>\n<\/tr>\n |
\n| ExploitGym<\/td>\n | 6.0<\/td>\n | 17.8<\/td>\n | 0.2<\/td>\n | 22.1<\/td>\n | 30.3<\/td>\n<\/tr>\n |
\n| ExploitBench<\/td>\n | 25.3<\/td>\n | 47.9<\/td>\n | 16.6<\/td>\n | 70.0<\/td>\n | 78.5<\/td>\n<\/tr>\n |
\n| SEC Bench Pro<\/td>\n | 47.5<\/td>\n | 66.3<\/td>\n | 17.7<\/td>\n | –<\/td>\n | 79.1<\/td>\n<\/tr>\n |
\n| Visual Agent<\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n<\/tr>\n |
\n| MiMo VisualCoding<\/td>\n | 71.5<\/td>\n | 72.3<\/td>\n | –<\/td>\n | 70.0<\/td>\n | 73.4<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n According to measurements by the publishers, MiMo-V2.6-Flash-RL scored 95.1 on CyberGym and 52.3 on AutomationBench v1.0.6, showing results that surpass other models in certain areas. It also achieved high-level scores of 87.6 on Terminal Bench 2.1, which measures whether commands can be actually executed in a terminal to complete tasks to the end, and 80.8 on OSWorld-Verified, which measures the ability to operate actual OS screens to complete tasks. On the other hand, it falls short of other large frontier models in cybersecurity fields such as ExploitBench (25.3) and ExploitGym (6.0), as well as Terminal Bench 4.0 (28.8), clearly dividing its strengths and weaknesses.<\/p>\n Strengths and Use Cases<\/h2>\nThe areas of strength and intended use cases for this model are based on the capabilities and training design of the original model, “XiaomiMiMo\/MiMo-V2.6-Flash-RL”.<\/p>\n \n- \n
Long-Context Processing with Integrated Multimodality<\/strong>
\n It supports native omni-modality, handling text, images, video, and audio within a single model. Furthermore, leveraging a context length of up to 1M tokens, it excels at analyzing vast software repositories as a whole, understanding lengthy tool execution histories (tool traces), or processing multi-session dialogues and agent operation logs without information loss.<\/p>\n<\/li>\n- \n
Autonomous Agent Operation and Environmental Adaptability<\/strong>
\n Through a batch reinforcement learning method called “You Only RL Once”, coding, general agents, visual tasks, and cybersecurity are trained in a complex manner within a single batch. This enables flexible adaptation and task execution in unfamiliar environments and tools, such as completing tasks via command execution in a terminal or visually recognizing and operating GUI screens of an OS.<\/p>\n<\/li>\n- \n
Fast Inference via Speculative Decoding<\/strong>
\n This model is equipped with a drafter for multi-token prediction (MTP) (utilizing a 5-layer SWA MTP structure) and supports speculative decoding to predict multiple tokens in a single forward pass. It is suitable for conversational assistants and agent executions where response generation speed and efficiency are required.<\/p>\n<\/li>\n<\/ol>\nThe main supported languages are English (en) and Chinese (zh), with expected use cases including conversational tasks, advanced tool calling, and complex multimodal analysis.<\/p>\n <\/p>\n Hardware Requirements<\/h2>\nEstimated requirements (calculated by Local Model Watch)<\/strong><\/p>\n |