MiMo V2.6 Flash vs MiMo V2.6 Pro

MiMo-V2.6-Flash is the efficiency-oriented checkpoint of Xiaomi's MiMo-V2.6 series, a natively omnimodal foundation model that accepts text, image, video, and audio in a single model and supports a one million token context window. The language backbone is a sparse mixture-of-experts transformer with roughly 309 billion total parameters and 15 billion activated per token, organized as 48 layers with 256 routed experts and top-8 routing. It uses a hybrid attention scheme that interleaves sliding window attention with global attention layers to cut key-value cache cost on long sequences, and pairs the backbone with a vision encoder, an audio encoder, and an audio tokenizer, plus a multi-token prediction module and a draft model for faster decoding.

Training emphasizes large scale reinforcement learning on verifiable, long-horizon tasks, with RL compute, environment diversity, and grader compute scaled together in a single mixed run. Xiaomi reports gains during RL on SWE-bench Verified, Terminal Bench, a cybersecurity benchmark, and an internal visual coding benchmark, reflecting a focus on agentic coding, computer use, and multimodal document and screen understanding rather than single turn chat.

MiMo V2.6 Pro is the flagship omni-modal foundation model in Xiaomi's MiMo V2.6 series, released as open weights alongside a Flash variant and a 9B distillation of Qwen3.5. It uses a sparse mixture-of-experts transformer with 1.02 trillion total parameters and roughly 42 billion activated per token, paired with a hybrid attention design that interleaves sliding-window and global attention layers to support a context window of about one million tokens. Dedicated encoders handle non-text inputs, including a vision encoder of roughly 681 million parameters and an audio tokenizer stack, so the model accepts text, images, video, and audio and returns text.

Post-training centers on large-scale reinforcement learning across thousands of interactive environments, combined with agentic grading, self-correction cold start, and a multi-prefix multi-teacher on-policy distillation stage that extends behavior to tasks that are hard to verify automatically. The resulting model targets long-horizon agentic work such as software engineering, terminal and computer-use operation, tool calling, cybersecurity analysis, and visual coding, and it reports gains over the prior MiMo generation on SWE-bench Verified, Terminal Bench, and internal visual coding and cyber benchmarks.