DeepSeek-R1 1.5B on SiMa.ai for less than 10 watts
Secure and private on-device deployment at the Edge with MLSoC Modalix.
SiMa.ai has announced the successful implementation of DeepSeek-R1-Distill-Qwen-1.5B on its ONE Platform for Edge AI, achieving breakthrough performance within an unprecedented power envelope of under 10 watts. This implementation marks a significant advancement in efficient, secure edge AI deployment.
The implementation is powered by the SiMa.ai ONE Platform with the MLSoC Modalix and Palette software, which demonstrates the exceptional responsiveness with Time to First Token (TTFT) as low as a few milliseconds, scaling efficiently based on query complexity. This performance metric positions SiMa.ai at the forefront of enabling conversational AI and multi-modal reasoning in real-world embedded edge AI applications where rapid response times are crucial.
“The availability of industry-leading open source models like DeepSeek R1 or Llama has democratized AI and accelerated the ability to deploy GenAI at the Edge,” said Krishna Rangasayee, CEO and founder of SiMa.ai. “SiMa.ai is leading this revolution in achieving this level of performance, completely at the edge, while maintaining a power envelope under 10W, opening new possibilities for secure, efficient AI deployment. SiMa.ai’s purpose-built MLSoC Modalix now enables multi-modal and high reasoning AI capabilities into diverse markets including robotics, automotive, medical, smart vision, and aerospace and defense applications.”
Details of this implementation include:
• DeepSeek-R1-Distill-Qwen-1.5B with industry-leading power efficiency at under 10W
• Rapid response times with TTFT ranging from 0.67 to 2.50 seconds with 32 input token size and 1028 output token size with room for optimization
• Optimization roadmap to extend performance to >30 Tokens per second (TPS)
Operating completely at the edge ensures that the user retains autonomy over their data while benefiting from high-performance AI capabilities at low power. This approach eliminates common security and privacy vulnerabilities typically seen with cloud deployments.
Organizations interested in learning more about the SiMa.ai ONE Platform and the DeepSeek R1 implementation can join the Modalix Early Access Program to explore our roadmap for LLMs and Generative AI solutions. To get started, visit https://sima.ai/modalix-eap/.
