San Jose-based SiMa.ai has announced the commercial rollout of its second-generation Modalix Machine Learning System-on-Chip (MLSoC), signaling a step forward in the era of Physical AI—where intelligence is embedded directly within devices. Designed to run under 10 watts of power, the MLSoC can now handle reasoning-based large language models (LLMs), transformers, convolutional neural networks (CNNs), and generative AI workloads directly at the edge.

The launch is accompanied by a System-on-Module (SoM) and supporting development kits. The 8GB SoM is priced at $349, while the 32GB variant costs $599, positioning the product as a commercially accessible option. According to SiMa.ai, the SoM is also pin-compatible with existing GPU modules, allowing companies to integrate it seamlessly into existing platforms for applications spanning robotics, automotive, healthcare, and industrial automation.

Edge AI with Low-Power Efficiency

Physical AI emphasizes instant, on-device decision-making in areas such as autonomous vehicles, smart cameras, and industrial robotics. Traditionally, running LLMs and advanced AI required power-hungry GPUs or cloud-based resources. By bringing this capability into a sub-10W power envelope, SiMa.ai addresses a major challenge: latency.

Processing data locally—whether from sensors, cameras, or text inputs—enables faster response times, reduces dependency on network connectivity, and enhances privacy. SiMa.ai claims its Modalix MLSoC delivers up to tenfold improvements in performance per watt compared to competing edge AI solutions. Its ARM-based architecture supports multiple workloads, combining efficiency with the ability to handle complex, multi-modal models in real time.

Software Stack for Developers

Alongside the hardware, SiMa.ai has introduced LLiMa, an automated compiler framework integrated into its Palette SDK. LLiMa simplifies the process of converting large language models and vision-language models (VLMs) into optimized binaries for the Modalix chip. This reduces deployment complexity, allowing developers to achieve high-throughput, low-power inference without extensive manual tuning.

With this framework, companies can extend AI capabilities to devices operating in constrained environments—delivering conversational, reasoning, and generative AI performance that was previously limited to high-power data centers. The addition of LLiMa further strengthens SiMa.ai’s ecosystem approach to Physical AI.

Benchmarking Practical Impact

To demonstrate real-world applicability, SiMa.ai showcased DeepSeek-R1 (1.5B) running on Modalix hardware. The results indicated fast Time-to-First Token (TTFT) responses and efficient throughput, all while staying below the 10W threshold. For industries where power efficiency is critical—such as drones, mobile robotics, and medical devices—this represents a meaningful advance.

The company argues that this level of performance opens up new opportunities for deploying AI at scale without the limitations of cloud connectivity or heavy infrastructure. It suggests Modalix can enable lightweight but powerful edge deployments that balance efficiency with intelligence.

Broader Market Implications

The release of Modalix reflects a broader movement within the AI industry: shifting from cloud-heavy architectures to on-device intelligence. While cloud-based training remains essential, the inference stage is increasingly moving closer to where data is generated. This not only reduces reliance on high-bandwidth connectivity but also ensures real-time responsiveness and better data security.

For enterprises, the implications are significant. Smart factories could deploy AI directly on assembly-line cameras; healthcare providers could run AI diagnostics at the patient’s bedside; and automotive companies could enhance safety features without relying on external networks. By combining energy efficiency, modularity, and compatibility, SiMa.ai is positioning its MLSoC as a competitive option for organizations building the next generation of AI-powered products.

Conclusion

SiMa.ai’s latest offering places emphasis on performance-per-watt as the defining metric for edge AI adoption. By shipping the Modalix MLSoC, SoM, and accompanying DevKits, the company is signaling its readiness to serve early adopters who need scalable and deployable AI solutions at the edge.

While the market for AI hardware remains competitive, SiMa.ai’s bet on Physical AI aligns with the growing industry consensus that intelligence must increasingly reside within devices themselves. As enterprises look for ways to deploy generative AI and reasoning-based models without overwhelming infrastructure or energy budgets, solutions like Modalix could play a pivotal role in shaping the edge AI landscape.