The Model Zoo built for edge AI.
Hundreds of pre-optimized models for benchmarking silicon, building proofs of concept, and proving edge AI is right for your product. The first model zoo built specifically for edge AI. Free, always.
AI models designed for constrained hardware.
Most AI model libraries and Zoo were designed for cloud inference, high-memory GPUs, or a single silicon ecosystem. Embedded teams working on constrained edge hardware need something different: models optimized for low-power devices, benchmarked on real silicon, and ready to evaluate quickly. ModelNova Zoo helps teams move from idea to proof of concept faster, without building deployment pipelines from scratch.
Built for constrained hardware
Every model is quantized, pruned, and compiled before it enters the Zoo. The starting point is a model that runs on a 512 KB MCU, not one adapted down from a cloud architecture.
Benchmarked on real silicon
Latency, RAM footprint, and accuracy numbers are measured by running the compiled binary on real hardware. Not estimated. Not simulated. Measured.
Faster proof-of-concept development
Pull a model, run it on your dev kit, and have real latency and memory numbers within hours. No deployment pipeline to build before you can evaluate.
Two products. One journey.
ModelNova has two products that work together. The Zoo takes you from idea to proof of concept, free models, real benchmarks, no setup. Fusion Studio takes you from proof of concept to finished product, retrain on your data, compile to your silicon, deploy. Zoo models load directly into Fusion Studio. No pipeline rebuild, no tool switch, no starting over.
From model selection to working proof of concept.
Move from idea to working proof of concept in days. Every model is quantized, benchmarked on real silicon, and ready to load into Fusion Studio when you’re ready to go further.
Optimized for constrained hardware
Models are quantized, pruned, and compiled before they enter the Zoo. The starting point is a model that runs on a 512 KB MCU — not one you have to adapt down from a cloud architecture.
Benchmarked on target silicon
Latency, RAM footprint, and accuracy numbers on every model card are measured by running the compiled binary on real hardware. Not estimated. Not simulated. Measured.
Everything needed for a working prototype
The Zoo includes pre-trained models, curated open-source datasets, and product blueprints. Everything in one catalog to take you from idea to working proof of concept without assembling tools from separate sources.
Ready to retrain in Fusion Studio
Zoo models are the start of your project, not the end. Load any model into Fusion Studio when you’re ready to retrain on your data, tune for your silicon, and compile to a production binary.
Models, datasets, and blueprints for embedded AI.
Computer Vision
The largest category in the Zoo. Object detection, segmentation, facial recognition, presence detection, people counting, gesture recognition, and visual anomaly detection — for camera-equipped edge hardware from microcontrollers to application processors.
Audio & Speech
Always-on keyword spotting, wake-word detection, audio denoising, speaker detection, and sound-event classification — running on MCU-class DSP or NPU hardware.
Sensing & Analytics
Datasets
A curated section of open-source datasets sourced and organized by our team. Useful for retraining a Zoo model on richer data, or for evaluating a model against your own accuracy benchmarks.
Blueprints
From idea to proof of concept then to production.
ModelNova Zoo helps teams evaluate models, benchmark silicon, and build working proof-of-concept applications quickly. Fusion Studio takes the next step: retraining, optimization, compilation, benchmarking, and deployment to production hardware.
Try before you download
Select Zoo models run live in your browser through the ModelNova Zoo’s ‘Try it Now’ feature. Use your device’s live camera feed in your web browser to try a model before you download it, so you know it’s the right model for you. The feature runs compiled WASM inference, not a cloud API call. What you see in the browser is the same model that you would download and deploy to silicon
Are you a silicon vendor?
Your developer ecosystem deserves its own branded model library. We port and optimize Zoo models for your hardware, package them in a co-branded portal, and operate it as the catalog evolves.
Four ways to start with the Zoo.
- Path 1
Validate silicon before committing to BOM
Your team is evaluating MCUs and NPUs for a new product. Pull a Zoo model that matches your workload, run it on candidate dev kits, and get honest latency and memory numbers within hours — before signing silicon orders.
- Path 2
Build a PoC for your next leadership review
You need to prove edge AI is the right path for your product line. Download a Zoo model, integrate it with the reference design, and have a working demo for your stakeholders in days — not months.
- Path 3
Find a starting point for a custom model
Off-the-shelf accuracy won't cut it for your final product, but you don't want to start from scratch. Load a Zoo model into Fusion Studio, retrain on your data, and ship a custom model in a fraction of the time.
- Path 4
Build something fun
Smart pet feeders. Plant-watering robots. Bird identifiers. Anything you've been waiting to build. The Zoo is free, always — bring your idea, pull a model, see what you can make happen.
Four ways to start with the Zoo.
Validate silicon before committing to BOM
Your team is evaluating MCUs and NPUs for a new product. Pull a Zoo model that matches your workload, run it on candidate dev kits, and get honest latency and memory numbers within hours — before signing silicon orders.
