Xiaomi Releases Open-Source AI Model to Challenge DeepSeek and Slash Inference Costs
Xiaomi launched MiMo-V2-Flash, an open-source AI model with 309 billion total parameters and 15 billion active parameters, marking the smartphone maker's aggressive push into artificial intelligence. The model matches performance of leading open-source competitors including DeepSeek-V3.2 and Kimi-K2 while delivering inference speeds of 150 tokens per second at costs of 0.1 per million input tokens and $0.3 per million output tokens.
The company released the model under MIT open-source license, with base version weights available on Hugging Face. MiMo-V2-Flash employs a mixture-of-experts architecture and supports deep thinking and internet search capabilities, targeting both conversational applications and real-time data scenarios.
Xiaomi's entry escalates competition in China's AI sector, where companies are racing to develop cost-effective alternatives to proprietary models. The release demonstrates how hardware manufacturers are building AI infrastructure to support their device ecosystems spanning smartphones, Internet of Things products, and electric vehicles.
The model is immediately available for free testing on Xiaomi's API Platform, with full inference code contributed to SGLang and complete technical documentation published on GitHub.
Benchmark Performance Rivals Closed-Source Models
MiMo-V2-Flash achieved top-two rankings among open-source models in AIME 2025 mathematics competition and GPQA-Diamond science knowledge testing. In coding capabilities, the model scored 73.4% on SWE-bench Verified, surpassing all open-source competitors and approaching GPT-5-High performance. The benchmark tests AI systems on fixing real-world software bugs, with the 73.4% success rate indicating practical programming problem-solving ability.
The model demonstrated 71.7% resolution rate on SWE-Bench Multilingual testing. In agent task evaluations using τ²-Bench, MiMo-V2-Flash scored 95.3 in communications, 79.5 in retail, and 66.0 in aviation categories. BrowseComp search agent scoring reached 45.4, jumping to 58.3 with context management enabled.
Writing quality approached top closed-source models, positioning MiMo-V2-Flash beyond tool functionality toward general-purpose assistance applications.
Technical Architecture Cuts Memory Requirements
Xiaomi implemented two core innovations to achieve cost reductions while maintaining long-context performance. The hybrid sliding window attention mechanism uses an aggressive 5-to-1 ratio alternating five layers of sliding window attention with one layer of global attention. The sliding window processes only 128 tokens, reducing KV cache storage by nearly six-fold while supporting context windows up to 256,000 tokens.
Chief scientist Luo Fuli noted on social media that window size 128 represents an optimal "sweet spot," with experiments showing that expanding to 512 tokens actually degrades performance. She emphasized that sink values are critical for maintaining performance and cannot be omitted during implementation.
The lightweight multi-token prediction module enables parallel prediction of multiple tokens, averaging 2.8 to 3.6 token acceptance and boosting inference speed by 2 to 2.6 times. The system functions during both inference and training phases, reducing GPU idle time. Luo acknowledged the team did not fully integrate MTP into reinforcement learning training cycles due to time constraints, but released the three-layer MTP implementation as open source.
Training Efficiency Breakthrough Reduces Compute by 98%
Pre-training utilized FP8 mixed precision across 27 trillion tokens with native support for 32,000-token sequences. The compression technique maintains accuracy while reducing memory usage and accelerating training, requiring deep optimization of underlying frameworks.
Post-training employed multi-teacher online policy distillation, a novel approach where the student model samples from its own policy distribution while multiple expert teachers provide dense reward signals at each token position. The method requires only one-fiftieth the computational resources of traditional supervised fine-tuning plus reinforcement learning pipelines while achieving teacher performance peaks.
MOPD supports flexible integration of new teachers, with student models capable of becoming teachers themselves to create self-reinforcing evolution cycles. Luo cited inspiration from Thinking Machine's on-policy distillation methodology, merging multiple reinforcement learning models to deliver efficiency gains that establish foundations for self-reinforcing loop systems.
For agent reinforcement learning scaling, Xiaomi's research team constructed over 100,000 verifiable tasks based on real GitHub issues. The automated pipeline runs on Kubernetes clusters with concurrent capacity exceeding 10,000 pods and 70% environment deployment success rates. Web development tasks utilize specialized multimodal validators that record video rather than static screenshots to verify code execution results.
Strategic AI Ecosystem Integration
The model integrates seamlessly with mainstream development environments including Claude Code, Cursor, and Cline, with the 256,000-token context window supporting hundreds of agent interaction rounds and tool calls. Xiaomi contributed all inference code to SGLang and shared optimization insights via LMSYS blog.
Technical reports disclose complete model specifications, with weights including MiMo-V2-Flash-Base published on Hugging Face under MIT license. The comprehensive open-source approach represents uncommon transparency among Chinese technology companies.
Luo indicated on social media that MiMo-V2-Flash represents "only the second step" in Xiaomi's AGI roadmap. The technical report acknowledges performance gaps compared to leading closed-source models, with plans to narrow differences through expanded model scale and training compute while exploring more robust agent architectures.
The release signals Xiaomi's strategic bet on AI infrastructure underpinning its hardware ecosystem spanning smartphones, IoT devices, and automobiles. The company appears to be replicating its 2011 smartphone disruption strategy, when it redefined flagship pricing at RMB 1,999, now applying similar cost-performance pressure to open-source AI models.