The global technology landscape has witnessed fresh momentum as officially launched its latest open-weight AI model, MiMo-V2-Flash, on December 16, 2025. Positioned as the most advanced model in Xiaomi’s MiMo family, MiMo-V2-Flash stands out not only for its speed and efficiency but also for its aggressive cost structure, directly challenging some of the biggest AI models currently dominating the market.
A Major Technical Leap in the AI Model Ecosystem
MiMo-V2-Flash features a massive 309 billion parameters, placing it among the largest AI models developed so far. However, its key innovation lies in its Mixture-of-Experts (MoE) architecture. Thanks to this design, only 15 billion parameters remain active during inference, allowing the model to deliver high performance while significantly reducing energy consumption and operational costs.
Xiaomi has released MiMo-V2-Flash as an open-weight model through its MiMo Studio developer portal, , and the company’s own API platform. This move enables developers worldwide to freely download, test, and integrate the model into their applications.
How Far Ahead Is Xiaomi on Speed and Pricing?
Performance metrics for MiMo-V2-Flash are drawing attention across the AI community. The model delivers an inference speed of 150 tokens per second, positioning it among the fastest AI models currently available.
Pricing has also been kept highly competitive. Xiaomi has set input token costs at $0.1 per million tokens and output tokens at $0.3 per million. By comparison, this makes MiMo-V2-Flash approximately 2.5–3.5% cheaper than premium models such as ’s Claude Sonnet 4.5, strengthening Xiaomi’s appeal for large-scale commercial use.
Benchmark Results: What Has MiMo-V2-Flash Achieved?
In benchmark testing, MiMo-V2-Flash has delivered standout results. On SWE-Bench Verified, the model achieved a 73.4% score, surpassing all other open-source AI models tested so far.
This performance places it slightly ahead of V3.2, which scored 73.1%, and very close to Claude Sonnet 4.5 at 77.2%.
Additional evaluations further highlight its strengths: 94.1% on AIME 2025, 95% on τ²-Bench Telecom, and 96% in agentic tool-usage tests. These results underline the model’s strong capabilities across reasoning, coding, and mathematical tasks.
Why MiMo-V2-Flash Directly Challenges DeepSeek and Approaches Claude
MiMo-V2-Flash is positioned as a direct competitor to DeepSeek V3.2. While both models show near-parity in reasoning benchmarks, Xiaomi’s model gains an edge in long-context evaluations.
On the Artificial Analysis Intelligence Index, MiMo-V2-Flash has scored 66, placing it on par with DeepSeek. Xiaomi attributes this advantage to the model’s 256K context window and a hybrid attention mechanism that combines local sliding-window attention with global attention in a 5:1 ratio. This architecture enables faster and more cost-efficient processing of long prompts.
What’s New in Xiaomi’s AI Architecture?
The strength of MiMo-V2-Flash lies in its engineering. Its Mixture-of-Experts design breaks large neural networks into specialized components, maintaining a balance between performance and efficiency.
Former DeepSeek researcher Luo Fuli, now part of Xiaomi’s MiMo team, described the model as the company’s second major step on its AGI roadmap. According to him, the newly developed hybrid Sliding Window Attention (SWA) architecture has outperformed other linear approaches during internal testing.
Xiaomi President Lu Weibing stated that the company’s AI investments are delivering results beyond expectations. He added that Xiaomi is now focused on deep AI integration into the physical world, including smartphones, tablets, and electric vehicles.
The Role of MiMo-V2-Flash in the Xiaomi Ecosystem
MiMo-V2-Flash is more than just another AI model—it is positioned as a foundation for Xiaomi’s future smart ecosystem. Designed as a general-purpose assistant, the model supports reasoning, coding, and agentic tasks.
Under Xiaomi’s “People-Car-Home” AIoT vision, the company plans to integrate MiMo-V2-Flash across smartphones, electric vehicles, and smart home devices, signaling a broader push to embed advanced AI capabilities throughout its product ecosystem.
Also Read
- Meta’s Big Bet: ‘Mango’ and ‘Avocado’ Accelerate the New AI Race
- Samsung Introduces Exynos 2600 — World’s First 2nm Smartphone Processor, Set to Debut in the Galaxy S26 Series
- Instagram Hashtags Rule: Only 5 Tags Allowed Now — Here’s How It Will Impact Creators and Brands
- Google Launches New Experimental AI Agent ‘CC’, Set to Transform Everyday Productivity
Utkarsh works as a Sub-Editor at 1Tak News. Technology and gadgets are his main beats, and he also tracks automobiles. A graduate of Mahatma Gandhi Kashi Vidyapeeth, he began his journalism career in 2023 and has built experience across digital media platforms like 1Tak. Besides technology, he also has considerable experience covering national affairs, politics, research, and international news.
