embedUR Systems spun off ModelNova as an independent company on July 16, 2026, targeting one of the hardest problems in edge AI: getting trained AI models out of the lab and into production on devices with less than 512 KB of RAM. Headquartered in San Ramon, California, with engineering operations in Chennai, India, ModelNova launches with nine semiconductor partners, an open-source model library, a desktop development environment called Fusion Studio, and NovaEyeD — a production-ready facial recognition pipeline already deployed in commercial products.
The spinoff comes from a company that has been bootstrapped for 22 years. embedUR Systems was founded in 2004 by Rajesh Subramaniam in Fremont, California, and has grown at an average of 35% per year with zero external investment. Now, ModelNova is betting that two decades of embedded software expertise can solve the problem that has kept edge AI stuck in the prototype phase: the gap between a working AI model and a shipping product.
Why Edge AI Deployment Is Hard
Edge AI deployment is the process of taking a trained machine learning model and running it locally on a physical device — a microcontroller, sensor, or embedded processor — rather than in the cloud. The reasons for doing this are straightforward: lower latency, reduced bandwidth costs, improved privacy, and the ability to operate without network connectivity.
The problem is that edge devices are severely constrained. A typical microcontroller used in TinyML applications has about 256 KB of RAM, runs at 48–80 MHz, and operates on less than 1 milliwatt of power. A standard AI model trained on a GPU cluster might be 2 MB or more. Fitting that model onto a device with a fraction of the memory requires aggressive optimization — quantization to shrink parameter sizes, pruning to remove redundant weights, and knowledge distillation to transfer knowledge from a large model to a small one — without losing meaningful accuracy.
But compression is only part of the challenge. Once a model fits in memory, it has to run on specific hardware. Every semiconductor vendor — ARM, STMicroelectronics, Infineon, NXP — has its own architecture, its own Neural Processing Unit (NPU), its own toolchain, and its own optimization requirements. A model that runs perfectly in a TensorFlow environment on a developer's laptop might need entirely different compilation and deployment steps for each target chip. Teams end up stitching together five separate tools: a training framework, a quantization script, a compiler, a benchmarking harness, and a deployment pipeline. Each tool has its own file formats, version dependencies, and failure modes.
Rajesh Subramaniam, Founder and CEO of embedUR Systems, framed the problem bluntly in ModelNova's launch announcement: "Why is it still so hard to turn a working AI model into a real product at the edge?"
What ModelNova Actually Ships
ModelNova's product portfolio targets three stages of the edge AI journey: starting from a pre-trained model, optimizing and validating it for target hardware, and shipping it in a commercial product.
ModelNova Zoo: Open-Source Starting Point
The ModelNova Zoo is a free, open-source library of over 200 pre-trained AI models, datasets, and application blueprints optimized for low-power devices. It covers vision, speech, and sound use cases. Developers can download a model, run it on a Raspberry Pi or on AI-native silicon from one of the nine semiconductor partners, and validate a concept in days rather than months.
The Zoo launched in June 2024 at the TinyML Summit in Milan and has since been integrated into the EDGE AI FOUNDATION's educational initiatives through EDGE AI LABS, enabling academic and industry collaboration on curated edge AI models.
Fusion Studio: Desktop IDE for Edge AI
Fusion Studio is a desktop application that brings model training, optimization, benchmarking, and deployment into a single environment. Instead of switching between five disconnected tools, a developer loads their data, selects or trains a model, quantizes it for the target hardware, benchmarks it for memory and latency, compiles it for the specific chip, and deploys it — all within one IDE.
A key feature of Fusion Studio is its local-first architecture. Training datasets, model weights, compiled binaries, and benchmark results never leave the developer's machine. The application connects to the internet only to download Zoo models that a user explicitly selects and to check for application updates (which can be disabled). For security, defense, and industrial teams working with sensitive data, this isn't a convenience — it's an architectural requirement driven by data sovereignty regulations.
Fusion Studio entered beta in October 2025 with Raspberry Pi support and expanded to the Arm Ethos-U NPU ecosystem in March 2026, starting with Alif Semiconductor's Ensemble microcontrollers (Ethos-U55 and Ethos-U85). The Explorer version is free to download with no time limit; Builder and Fleet tiers add advanced AutoML, on-prem hosting, and dedicated tooling.
Fusion Studio currently supports:
- Arm Ethos-U85 (NPU) — Production status, toolchains: ExecuTorch, Vela, TFLite
- Arm Ethos-U55 (NPU) — Production status, toolchains: ExecuTorch, Vela, TFLite
- Alif Ensemble E3/E7/E8 (Ethos-U55/U85 + Cortex-M55) — Production status
- STMicroelectronics STM32N6 (Neural-ART Accelerator) — Beta status, toolchain: TFLite
- Infineon PSoC Edge E84 (Ethos-U55 + Cortex-M55) — Beta status, toolchains: TFLite, ExecuTorch
- Raspberry Pi 4/5 — Prototyping and demos
NovaEyeD: Facial Recognition in Production
NovaEyeD is ModelNova's commercial face recognition pipeline — and the proof point that the company ships production AI, not just prototypes. It runs entirely on-device: detection, liveness detection, enrollment, and 1:N face matching, all within MCU-class memory budgets of less than 20 MB of SRAM.
The system processes face recognition in approximately 40 milliseconds per frame — about 10 ms for face detection and 30 ms for recognition — fast enough for real-time, continuous recognition at entry points. It operates fully offline with zero network round-trips, meaning biometric data never leaves the device. This architecture sidesteps the compliance gap that disqualifies cloud-based face recognition from regulated deployments under frameworks like the EU AI Act and India's Digital Personal Data Protection (DPDP) Act.
NovaEyeD is already deployed in commercial products, including home appliances and sensing applications. It has been validated on production silicon from STMicroelectronics (STM32N6), Infineon, Synaptics, and Alif Semiconductor. STMicroelectronics features NovaEyeD as a production reference for facial recognition on STM32N6 silicon in its Edge AI Suite case study library, and the two companies co-hosted a technical webinar demonstrating the pipeline.
Building production-grade face recognition from scratch — including detection, liveness, alignment, feature extraction, 1:N matching, and drift handling — typically takes embedded teams 18 or more months to reach commercial-grade accuracy. NovaEyeD's license model delivers the entire pipeline optimized for a customer's chosen silicon, with a reference application, an enrollment SDK, and ongoing engineering support.
Nine Semiconductor Partners and a Foundation Leadership Role
ModelNova's semiconductor partnerships are central to its value proposition. The company works with nine partners across the edge AI silicon ecosystem:
- Arm Holdings — Architecture partner, Ethos-U NPU ecosystem
- STMicroelectronics — STM32N6 with Neural-ART Accelerator
- Infineon Technologies — PSoC Edge E84
- Alif Semiconductor — Ensemble Series (E3/E7/E8)
- NXP Semiconductors
- Silicon Labs
- Renesas
- Synaptics
- Ceva
Each partnership is a co-development relationship, not a logo on a webpage. ModelNova validates every integration on real hardware — not simulation — and the partners publicly endorse the collaboration. Reza Kazerounian, Co-Founder and President of Alif Semiconductor, confirmed that ModelNova has been "the partner proving what's possible" on Alif's Ensemble MCU architecture, "from free Zoo models to production-ready deployments optimized for our MCUs."
ModelNova is also a Leadership Partner of the EDGE AI FOUNDATION (formerly the tinyML Foundation), the global community advancing edge AI research, education, and adoption. Pete Bernard, CEO of the EDGE AI FOUNDATION, called ModelNova's independence "exactly the kind of maturation we hoped to see" from the ecosystem.
The Chennai Engineering Bet
ModelNova's engineering operations are anchored in Chennai, India, within the embedUR organization. The Chennai office was established in 2008 and now houses over 400 engineers — the majority of the company's technical team. embedUR has committed Rs 500 crore (approximately $60 million USD) in investment in Tamil Nadu over five years, with 2026 representing the third phase of this expansion.
A dedicated Edge Intelligence Center with a 100-engineer capacity was established to power the development and deployment of on-device intelligence across telecom, industrial, and semiconductor industries. The company plans to increase headcount from 400 to 550+ engineers by the end of 2026, hiring across embedded engineering, AI/ML, data science, silicon specialization, and signal processing.
Dr. T.R.B. Rajaa, Minister for Industries, Investment Promotions and Commerce in the Government of Tamil Nadu, stated that "embedUR's investment in expanding its R&D and Edge Intelligence capabilities in Chennai aligns closely with our focus on tech product development" and reinforces "Tamil Nadu's position as one of India's leading tech development hubs."
The Chennai center's significance goes beyond cost. ModelNova's about page emphasizes that the company designs within real hardware constraints — 512 KB of RAM, 0.1 W of power, GOPS-class NPUs — every day. The engineering team's 22 years of experience shipping software on constrained hardware is the foundation that lets ModelNova optimize models for specific silicon rather than adapting cloud AI frameworks downward.
How ModelNova Compares to the Alternatives
The edge AI tooling landscape is fragmented. Most solutions address one stage of the pipeline.
Cloud-originated frameworks (TensorFlow Lite, PyTorch Mobile) provide model conversion and basic runtime support but leave the optimization, benchmarking, and deployment workflow to the developer. Teams still need to stitch together separate tools for quantization, compilation, and hardware-specific deployment.
Silicon vendor tools (Arm Vela, ST Edge AI Suite, Infineon's tools) are optimized for their own chips but create vendor lock-in. A team deploying across multiple silicon platforms ends up learning and maintaining multiple vendor-specific toolchains — exactly the fragmentation problem ModelNova was built to solve.
AutoML platforms attempt to hide the complexity behind automated model selection and optimization. ModelNova's position is that embedded engineers don't want a magic box that guesses at their architecture and quantization strategy — they want a real IDE that exposes every step of the pipeline and gives them control over every decision, with AutoML available as a tool, not as the product.
ModelNova's differentiation is the combination of all three: an open-source model library for starting, a hardware-agnostic IDE for building and deploying, and production-ready licensable models for teams that need to ship immediately rather than build from scratch. The 400-engineer embedded team provides the customization and integration support that pure-product vendors cannot offer — as the company puts it, "Pure-product vendors point you at their roadmap. We point you at our engineering team."
Market Context: Why This Matters Now
The edge AI market is growing rapidly, driven by demand for real-time data processing, reduced latency, enhanced privacy, lower bandwidth usage, and the proliferation of IoT devices. Market research firms project significant growth, though estimates vary:
- Grand View Research estimates the edge AI market at $24.9 billion in 2025, projected to reach $118.7 billion by 2033 at a CAGR of 21.7%
- BCC Research estimates the market at $8.7 billion in 2024, projecting $56.8 billion by 2030 at a CAGR of 36.9%
- IndustryARC projects $110 billion by 2031 at a CAGR of 29.5%
- Polaris Market Research estimates $21.58 billion in 2025, growing to $108.57 billion by 2034 at a CAGR of 19.7%
Regardless of which forecast is closest, the trajectory is clear: edge AI is moving from experimentation to production at scale. North America dominated the market with a 34.5% share in 2025, but Asia Pacific is expected to witness the fastest growth, driven by smart city and automation projects.
The companies that win in this market will be the ones that solve the deployment problem — not the model-building problem. Model-training infrastructure is mature. Cloud-based development environments are abundant. The bottleneck is getting a working model compiled, optimized, benchmarked, and deployed on a specific piece of silicon within the memory and power budget of a real product.
That's the gap ModelNova is built to close. Whether it succeeds will depend on execution: whether Fusion Studio can expand beyond its current supported platforms fast enough to cover the long tail of edge hardware, whether the open-source Zoo attracts enough developer adoption to build network effects, and whether the licensable model business (NovaEyeD and future products) can sustain the company while the platform tools mature.
The Founder's Bet
Rajesh Subramaniam founded embedUR Systems in 2004 after a decade at companies including Schlumberger, Bay Networks, Atoga Systems, and ARRIS. He holds a bachelor's degree from Annamalai University in Tamil Nadu and a master's in Electrical Engineering from the University of Arkansas. He built embedUR into a profitable, bootstrapped company with Fortune 500 customers and 35% average annual growth — without taking a single dollar of external investment.
Now he's spinning off the company's edge AI business as an independent entity with its own brand, roadmap, and commercial identity. In his words: "ModelNova began as our answer to a simple question: why is it still so hard to turn a working AI model into a real product at the edge? Establishing ModelNova as an independent company lets us go after that problem with full focus — giving developers a free place to start, a platform to build on, and production-ready models they can license and ship."
The bet is that 22 years of embedded engineering experience, combined with an open-source-plus-commercial-platform-plus-licensable-models business model, can outcompete both cloud-first AI tooling companies that don't understand hardware constraints and silicon vendors whose tools only work on their own chips.
For developers and engineering teams building edge AI products, ModelNova's launch means there is now a credible, hardware-agnostic platform with a free starting point, a production-grade IDE, and shipping commercial models — backed by a company that has been solving embedded problems since before edge AI had the name.
Discussion
0 comments