SiMa.ai launches chip enabling complex LLMs on low-power edge devices.

In a significant move to advance artificial intelligence beyond the cloud, AI chip company SiMa.ai has commenced production of its next-generation platform, designed to run complex, reasoning-based large language models (LLMs) on edge devices while consuming minimal power. The company has launched its MLSoC Modalix, a second-generation Machine Learning System-on-Chip, which aims to power the burgeoning field of "Physical AI." This category of AI involves embedding intelligence directly into physical systems like robots, industrial machinery, autonomous vehicles, and healthcare devices, enabling them to perceive, interact, and learn from their environments without constant reliance on remote data centers.[1][2] The new technology is engineered to operate under 10 watts, a critical threshold for devices where power efficiency is paramount.[3][4] This launch represents a pivotal step in making sophisticated, multi-modal AI—capable of processing text, images, and other data types simultaneously—a practical reality at the edge.

The core of the new offering is the MLSoC Modalix chip, which is being delivered to customers via a System-on-Module (SoM) and an accompanying development kit.[1][3] This hardware is purpose-built to handle a diverse range of AI workloads, a crucial feature for the evolving needs of the edge market.[1] Unlike traditional processors, the Modalix platform is not limited to a single type of AI model; it can efficiently run large language models like Meta's Llama, as well as vision language models (VLMs), transformer models, and conventional convolutional neural networks (CNNs).[1] This flexibility is vital for creating devices that can perform complex reasoning and make real-time decisions based on multiple inputs.[1][4] To facilitate adoption, the Modalix SoM is designed to be pin-compatible with leading Graphics Processing Unit (GPU) SoMs, allowing developers to integrate the new chip into existing systems as a drop-in replacement, significantly reducing development time and costs.[3][5] The company boasts that its platform can deliver a tenfold increase in performance and energy efficiency compared to alternative solutions.[6][7] For instance, the 50 TOPS (trillion operations per second) version of the chip can run the Llama2-7B model at a speed of more than 10 tokens per second, a notable performance metric for a low-power edge device.[8]

SiMa.ai has consistently emphasized its identity as a software-centric company, and this launch reinforces that strategy.[9][10] The hardware is supported by a robust software ecosystem designed to simplify the development process, which has historically been a major barrier to entry in the embedded edge market.[9] A key component of this is the Palette Edgematic software, a no-code, graphical user interface that allows developers to create and deploy machine learning pipelines by dragging and dropping elements on a visual canvas.[11][9] This "push-button" approach is intended to make ML development more accessible to a wider range of engineers and companies.[6][9] Complementing this is the newly introduced LLiMa software framework, a unified system for running LLMs, LMMs (Large Multimodal Models), and VLMs on the Modalix chip with zero cloud dependency.[1][5] LLiMa allows developers to import open-source or custom models, which are then automatically optimized into binaries ready to run on the device, enabling fully on-device Retrieval-Augmented Generation (RAG) and other advanced AI capabilities.[5]

The implications of enabling complex, reasoning-based AI on low-power devices are far-reaching. By processing data locally, systems can achieve the low latency necessary for safety-critical and time-sensitive applications.[1] In industrial automation, robots can perform more nuanced tasks; in automotive, in-car systems can offer more intelligent and responsive infotainment and driver assistance; and in healthcare, medical imaging devices can provide quicker, more detailed analysis at the point of care.[1][2][4] This represents a fundamental shift from edge devices that primarily perform simple computer vision tasks, like object detection, to systems that can engage in visual reasoning—understanding the context of a scene and interacting with it in a more human-like way.[1] This move towards "Physical AI" is about creating smarter, more autonomous systems that are not tethered to a network connection for their core intelligence.[1] With its Modalix SoM and DevKit now available, priced for enterprise deployments starting at $349 for an 8GB module, SiMa.ai is positioning itself to be a key enabler of this next wave of AI-driven innovation.[1][3]