MiMo-V2-Flash

Another major feature group is its ultra-long context and agentic workflow support, centered on a 256k token context window. This extensive memory allows the model to maintain coherence and relevance across hundreds of rounds of agent interactions and tool calls, enabling complex, multi-step tasks like automated software debugging, data analysis pipelines, and interactive web development. The model can generate functional HTML webpages with one click, working seamlessly with vibe-coding scaffolds such as Claude Code, Cursor, and Cline, as showcased by its ability to create a fully interactive macOS simulation with a working terminal, file manager, and system settings from a single prompt. This long-context capability, combined with the hybrid sliding-window attention architecture, ensures efficient processing of lengthy documents, codebases, and conversational histories without performance degradation, which is critical for sustained agentic deployments and enterprise applications.

The model incorporates Multi-Token Prediction as a native draft model for self-speculative decoding, delivering real deployment speedups by generating multiple draft tokens for the main model to verify in parallel. This approach lifts the arithmetic intensity of both FFN and attention computations without increasing KV cache I/O, addressing the memory-bound nature of LLM decoding. In MiMo-V2-Flash, the MTP block uses a dense FFN to limit parameters and SWA to reduce KV cache and attention costs, achieving an accepted length of 2.8–3.6 tokens and an effective speedup of 2.0–2.6×. Additionally, the model supports a post-training paradigm called Multi-Teacher Online Policy Distillation, which efficiently scales reinforcement learning to enhance reasoning and agentic capabilities using token-level rewards from multiple teachers, requiring less than 1/50 of the computational resources of traditional SFT+RL pipelines.

The overall workflow of MiMo-V2-Flash is designed for high-throughput inference and seamless integration into developer environments. It leverages a 1:5 hybrid of Global Attention and Sliding Window Attention, with empirical results showing SWA delivers better overall performance than Linear Attention across general tasks, long-context payloads, and reasoning, while providing a fixed-size KV cache for easy integration with existing training and inference infrastructure. The model's parallel decoding, enhanced by MTP, enables extremely high output token throughput. Users can interact via its API platform, AI Studio chat interface, or open-source weights on Hugging Face, employing it for everything from instant code generation to prolonged agentic sessions. The decoupled design of MOPD supports flexible integration of new teachers and ORMs, enabling a 'teach and learn' closed-loop for continuous self-improvement.

Concrete use cases include software engineering teams utilizing the model to automatically resolve GitHub issues, as evidenced by its SWE-bench scores, where it can understand bug reports, navigate code repositories, and generate correct patches. Developers can leverage its one-click HTML generation for rapid prototyping of web applications, such as creating a macOS simulator with functional terminal commands and file management. Researchers employ it for complex mathematical and scientific reasoning, tackling problems from the AIME competition or GPQA diamond-tier questions. Agentic scenarios involve deploying it as an autonomous assistant that performs hundreds of tool calls across extended sessions, like managing cloud infrastructure or conducting multi-step data analysis. The outcomes are faster development cycles, reduced manual coding effort, and the ability to automate intricate, knowledge-intensive tasks with high accuracy and cost efficiency.

The target audience includes software engineers, AI researchers, data scientists, and enterprises needing a high-performance, open-source model for coding, reasoning, and agentic applications. It is particularly suited for platforms and tools like Claude Code, Cursor, and Cline for vibe-coding, and integrates with inference frameworks like SGLang, to which Xiaomi contributed all inference code on Day 0. The tech stack involves the model's hybrid attention architecture, MTP for speed, and MOPD for efficient post-training, all available under the MIT license on Hugging Face. Pricing is structured at $0.1 per million input tokens and $0.3 per million output tokens, with a limited-time free tier on the API Platform. In summary, MiMo-V2-Flash delivers top-tier intelligence with unprecedented speed and affordability, making advanced AI reasoning and coding accessible for both open innovation and commercial production.