News
Jan 7, 2026
News
Open-Source
Artificial Intelligence
Machine Learning
NewDecoded
3 min read
Image by xiaomiforall
Xiaomi has officially launched MiMo-V2-Flash, a new open-source foundation language model that sets a high bar for speed and reasoning. Released on December 16, 2025, the model is designed to excel in coding and agentic tasks while remaining accessible for everyday assistance. It is currently available for global use through Hugging Face and Xiaomi’s specialized API Platform.
This 309 billion parameter model uses a Mixture-of-Experts architecture that only activates 15 billion parameters at a time. This design choice enables a lightning-fast inference speed of 150 tokens per second, which is a significant leap for models of this scale. By utilizing a hybrid attention mechanism, MiMo-V2-Flash maintains a deep 256k context window without the memory constraints typical of full-attention models.
Performance benchmarks place MiMo-V2-Flash at the top of the open-source field, particularly in software engineering. It currently holds the number one spot on the SWE-bench Verified and Multilingual leaderboards, outperforming notable competitors like DeepSeek-V3.2. Its reasoning capabilities are equally impressive, rivaling closed-source giants like GPT-5 in the AIME 2025 mathematics competition.
Efficiency is the defining characteristic of this release, as Xiaomi has priced the model at just $0.10 per million input tokens. The introduction of Multi-Token Prediction technology allows the system to draft and verify multiple tokens in parallel, resulting in a speedup of over two times. This makes high-end AI intelligence more affordable and scalable for developers and enterprise users alike.
Engineered for agent-first scenarios, the model supports functional HTML generation and integrates smoothly with tools like Cursor. It can handle hundreds of rounds of tool calls without losing focus, making it a strong candidate for autonomous software development. Detailed model specifications can be found in the technical report released alongside the weights. This release marks a significant milestone for Xiaomi as it builds a leadership position in the open-source AI ecosystem. By providing high-intelligence models at minimal cost, the company is making its hardware and software a primary destination for AI development. This strategy aims to challenge the dominance of major AI labs by making frontier-level capabilities accessible to everyone.
The arrival of MiMo-V2-Flash signals a new phase where efficiency and open access are the primary drivers of innovation. As proprietary labs focus on massive scaling, Xiaomi is proving that architectural breakthroughs can deliver comparable intelligence at a fraction of the cost. This move will likely force a market-wide shift in pricing and push the industry toward a future where high-end reasoning is a standard commodity.