In a significant move to solidify its position in the competitive artificial intelligence landscape, Chinese tech giant Xiaomi has officially launched and open-sourced its latest large language model, the MiMo-V2-Flash. The announcement, made at the 2025 Xiaomi "Human-Car-Home Full Ecosystem" Partners Conference on December 17, marks a strategic push to embed advanced AI capabilities across its vast hardware portfolio, from smartphones to smart homes and electric vehicles. Leading this charge is the newly appointed head of Xiaomi's MiMo model team, the highly-regarded AI researcher Luo Fuli, whose public debut underscored the company's serious commitment to developing foundational AI technology.
The MiMo-V2-Flash: Engineered for Speed and Efficiency
The MiMo-V2-Flash is a Mixture of Experts (MoE) model with a total parameter count of 309 billion, of which 15 billion are active during inference. This architecture choice is central to its design philosophy, which prioritizes high-speed performance and cost-effectiveness over sheer scale. Xiaomi's team, led by Luo Fuli, focused on optimizing the model specifically as a base for AI agents, emphasizing capabilities like code generation and tool use that are essential for interactive systems. The company claims the model achieves a remarkable inference speed, generating up to 150 tokens per second, which is 2 to 2.6 times faster than some comparable models. This speed is achieved through technical innovations like a three-layer MTP inference acceleration for parallel token verification and a hybrid attention mechanism combining sliding window and global attention.
MiMo-V2-Flash Key Specifications
- Model Type: Mixture of Experts (MoE)
- Total Parameters: 309 Billion (309B)
- Active Parameters: 15 Billion (15B)
- Claimed Inference Speed: ~150 tokens/second
- Speed Claim vs. Benchmark: 2.0x - 2.6x faster than unspecified comparable models.
- API Pricing: Input: CNY 0.7 / 1M tokens; Output: CNY 2.1 / 1M tokens.
- Notable Benchmark Score: 73.4% on SWE-Bench Multilingual.
- Primary Focus: Code generation, tool use, and serving as a base for AI agents.
Performance and Benchmark Claims
In terms of raw performance, Xiaomi positions the MiMo-V2-Flash as a top-tier contender among open-source models. Luo Fuli stated that in code and agent capability benchmarks, the model ranks within the global top two for open-source offerings. Specifically, it reportedly scored 73.4% on the SWE-Bench Multilingual software engineering benchmark, a result that Xiaomi claims surpasses several leading closed-source models. The company provided direct comparisons, noting that while its performance is on par with or exceeds models like DeepSeek-V3.2 and Kimi K2-Thinking in many evaluations, it does so with only one-half to one-third of their parameters. This "smaller but smarter" approach is a key part of Xiaomi's narrative, suggesting a move away from pure scaling laws toward more efficient, specialized model training.
Reported Performance Comparisons
- Vs. Claude Sonnet 4.5: Claimed inference cost is 2.5% of Claude's; speed is 2x faster.
- Vs. DeepSeek-V3.2 / Kimi K2-Thinking: Claims comparable or superior performance on many benchmarks with 1/2 to 1/3 of the parameters.
- Positioning: Ranked as Top 1-2 in global open-source models for code and agent capabilities according to Xiaomi.
A Disruptive Pricing Strategy
Perhaps the most aggressive aspect of the launch is the model's pricing. Xiaomi has set the API cost for MiMo-V2-Flash at CNY 0.7 per million tokens for input and CNY 2.1 per million tokens for output. The company framed this as a major competitive advantage, asserting that the inference cost is merely 2.5% of a benchmark model like Claude Sonnet 4.5. This aggressive pricing strategy echoes Xiaomi's historic playbook of disrupting markets with high-value, low-cost products, now applied to the cloud AI inference sector. The goal is clearly to lower the barrier to entry for developers and businesses looking to build agentic AI systems, potentially accelerating adoption within Xiaomi's own ecosystem and beyond.
The Strategic Vision: AI as the "Brain" for Xiaomi's Ecosystem
The launch was not an isolated event but a core component of Xiaomi's broader "Human-Car-Home Full Ecosystem" strategy. Luo Fuli's presentation emphasized that the next generation of AI should not be a mere "language simulator" but an intelligent agent that understands and interacts with the physical world. This vision directly aligns with Xiaomi's needs. The model is intended to serve as the AI foundation for upgrading its voice assistant "Xiao Ai," the HyperOS, and, critically, its autonomous driving systems. By developing an in-house, efficient base model, Xiaomi aims to create a cohesive AI experience across its hundreds of millions of connected devices, from phones and appliances to the Xiaomi SU7 and YU7 electric vehicles.
Leadership and Future Roadmap
The spotlight on Luo Fuli, often dubbed the "AI genius少女" (prodigy) in Chinese media, highlights Xiaomi's investment in top talent to achieve its AI ambitions. Having joined Xiaomi just over a month ago after notable tenures at Alibaba's Damo Academy and DeepSeek's parent company, her rapid debut signifies the project's high priority. Xiaomi has committed substantial resources to this endeavor, with plans to invest over CNY 200 billion in R&D over the next five years, a significant portion dedicated to AI. The MiMo-V2-Flash is the latest in a series of model releases in 2025, following others focused on vision and embodied intelligence, indicating a rapid and structured development pace.
Context: Xiaomi's 2025 AI Model Releases
- April 2025: Open-sourced MiMo-7B series (base, instruction-tuned, RL versions).
- May 2025: Released MiMo-VL-7B, a multimodal vision-language model.
- November 2025: Introduced MiMo-Embodied, a model for autonomous driving and robotics.
- December 2025: Launched flagship MiMo-V2-Flash, focused on agentic AI and efficiency.
Conclusion: A Calculated Bet on AI's Future
Xiaomi's launch of the MiMo-V2-Flash represents a calculated and multifaceted strategic bet. It is a technical demonstration of efficient model design, a market play with disruptive pricing, and a foundational step for its entire product ecosystem. While the model's real-world performance in diverse applications remains to be thoroughly validated by the developer community, its specifications and benchmarks present a compelling case. For Xiaomi, success will ultimately be measured not by leaderboard positions, but by how seamlessly and powerfully this AI "brain" can be integrated into the daily experiences of its users, powering everything from a smarter phone to a safer, more intelligent car. The race to define the next era of consumer AI is on, and Xiaomi has just unveiled a potent new contender.