SiMa.ai Modalix MLSoC and Palette Software

SiMa.ai is a San Jose, California based company founded in 2018 that builds machine learning systems-on-chip and software for physical AI. It targets robots, drones, industrial machines, automotive systems, smart vision and healthcare devices that need to run vision, language and generative AI models locally within a tight power budget.

The Modalix MLSoC family combines a machine learning accelerator with eight Arm Cortex-A65 CPUs, an image signal processor, vision and DSP blocks, 4 x 10G Ethernet and PCIe Gen5 on TSMC N6, in configurations from 25 to 200 TOPS; the main Modalix part delivers 50 TOPS at under 10 W and ships as a chip, system-on-module and devkit. The Palette software suite compiles and deploys CNN, transformer and LLM workloads across the family, and Palette Neat adds an agentic development environment. Named customers and partners include Bosch, Emerson, TRUMPF, Kontron, Micron, Synopsys and AVerMedia.

• Modalix MLSoC: 50 TOPS at under 10 W, with BF16 and INT8 support for CNN, ViT and LLM workloads.
• Scalable family: M25 to M200 configurations sharing one software stack.
• Integrated SoC: eight Arm Cortex-A65 CPUs, ISP with MIPI CSI-2, 4 x 10G Ethernet and PCIe Gen5.
• Palette software: compiler and deployment suite for heterogeneous compute on the device.
• Modalix DevKit: runs a full ML pipeline on one system-on-module, including on-device conversational AI.
• Palette Neat: agentic development environment to shorten projects from months to days.

An industrial machine builder in Vietnam can integrate Modalix modules into inspection or packaging equipment for on-board vision. A robotics or drone startup can run perception and voice interaction locally on one chip. A camera maker or system integrator can build video analytics boxes that summarise scenes with vision-language models without sending video to the cloud.

Runs multimodal and generative AI on-device at under 10 W, removing cloud latency and data transfer for robots and cameras. One software suite covers the whole MLSoC family, so models move between 25 and 200 TOPS parts without rework. The company claims about 10 times the performance per watt of alternatives. Integrated ISP, CPUs and 10G Ethernet reduce the number of chips needed in a vision system.

Humanoid robots, drones, ADAS and smart cameras are moving generative and multimodal AI onto the device for latency, privacy and connectivity reasons. Manufacturers in Asia that build robots, industrial PCs and cameras need efficient silicon and toolchains, and Southeast Asian factories adopting automation and machine vision are a growing market for such edge platforms.

Founded in 2018, with press releases datelined San Jose, California. Raised a 150 million USD Series C in September 2026 co-led by Fidelity and Amplify, reaching 500 million USD in total funding and a 1.45 billion USD valuation. Reports that revenue quadrupled from 2024 to 2025. Partners include Wobot.ai for reasoning-based video analytics, and Micron is a strategic investor.