Frequently Asked Questions

What does ModelCat actually do?

ModelCat is an autonomous AI system that builds, optimizes, and validates custom machine-learning models for your target hardware.

Instead of manually selecting neural network architectures, tuning methods, and hoping the model will work in production, ModelCat uses proprietary AI to explore the full solution space, generating thousands of possible solutions and testing them against real-world constraints defined up front by the user: chip power, latency, memory, and behavior.

Simply put: ModelCat’s AI builds production-ready AI that works in the real world.

Who is ModelCat for?

ModelCat is designed for teams creating AI features that must run reliably in production environments, especially on edge and embedded hardware.

Typical users include:

Technical leaders evaluating AI deployment and platform strategy
Product and platform teams building AI-powered features
Embedded and systems engineers deploying models to hardware
AI practitioners and data scientists working with real-world deployment constraints

You don’t need to be an AI research expert to use ModelCat. It’s designed to be usable by product and systems teams, not just AI specialists.

At the same time, experienced AI practitioners can access ModelCat’s Advanced Mode, which provides deeper control over architecture choices, optimization strategies, and model design decisions.

What types of AI workloads does ModelCat support?

ModelCat is well-suited for edge and embedded AI workloads, including:

Computer vision
Time-series and sensor data
Event detection and classification
Domain-specific (including narrow, production-critical use cases, even if the use case is outside the standard ML workloads)

The system is optimized for scenarios where reliability, efficiency, and deployment confidence matter more than abstract benchmark scores.

What neural network architectures does ModelCat support?

ModelCat does not rely on a fixed library of predefined neural network architectures.

Instead, ModelCat’s AI explores the overall architecture design space and generates architectures tailored to your specific use case, dataset, and hardware constraints. This allows the system to produce models optimized for your specific requirements, rather than forcing your problem into a predetermined model structure.

ModelCat has already explored and trained on more than 50,000 neural network architectures and other architectures, giving the system access to a design space that is orders of magnitude larger than what a typical engineering team could evaluate manually. This allows ModelCat’s AI to discover novel solutions that are out of reach to human-driven workflows.

Depending on the task, ModelCat can generate architectures inspired by modern deep learning techniques such as convolutional networks, transformers, hybrid architectures, and other emerging designs.

ModelCat also incorporates advances from modern AI research, allowing the system to automatically apply state-of-the-art architectural ideas.

How does ModelCat improve my models over time?

ModelCat improves continuously as it builds models across different datasets, hardware targets, and deployment constraints—forming an ongoing learning loop that strengthens future model generation.

Each completed model contributes new information about architecture design, optimization strategies, and real hardware behavior. Over time, this allows ModelCat’s AI to explore the solution space more effectively and identify stronger architectures faster.

In addition to learning from its own builds, ModelCat’s AI is regularly updated with advances from academic research and industry best practices. This allows the system to combine cutting-edge ideas with practical deployment experience, helping teams build production-ready models more quickly and reliably.

What kinds of models can ModelCat generate?

ModelCat can generate models tailored to a wide range of machine learning tasks and deployment scenarios.

Since ModelCat’s AI designs models based on the problem, dataset, and hardware constraints, it can produce architectures suited to tasks such as:

Computer vision
Physical AI Systems
Sensor and time-series analysis
Event detection and classification
Embedded and edge AI workloads

Rather than selecting from a predefined set of models, ModelCat generates architectures optimized for the specific deployment environment. This allows the resulting models to balance accuracy, efficiency, and hardware compatibility in ways traditional workflows often cannot.

How is ModelCat different from traditional ML workflows?

Traditional ML workflows require teams to:

Manually choose neural network architectures
Tune models through trial and error
Validate on real hardware late in the development process — often after months of work
Discover hardware constraints too late to fix them inexpensively

ModelCat flips that process.

With ModelCat:

Constraints are inputs, not surprises
Models are built specifically for your inference-target hardware
Performance is validated on real hardware
Time to deployment is measured in days, not months

How does ModelCat stay current with the latest AI research?

The field of AI evolves rapidly, with new architectures, training strategies, and optimization techniques emerging constantly.

ModelCat is designed to continuously incorporate advances from academic research and industry best practices. Our team regularly trains ModelCat’s AI on new architectural ideas, training methods, and optimization strategies so the system can evaluate and apply them automatically when building models.

This means teams benefit from state-of-the-art techniques without needing to monitor research papers, evaluate new repositories, or manually integrate emerging methods into their workflows.

ModelCat effectively acts as a bridge between cutting-edge research and practical production deployment.

Does ModelCat support language models?

Not yet, but we are definitely working on it. If you are interested in learning more, contact us.

How long does it take to get a model?

In most cases, ModelCat can deliver optimized, ready-to-run models in 24-48 hours or less, depending on dataset size and AI task. Our AI-powered approach avoids the long iteration cycles of manual workflows.

Can I control model-building decisions?

Yes. ModelCat accelerates execution, but it does not remove control. By default, you define:

Target hardware
Performance constraints
Optimization priorities (accuracy, speed, power, memory)
Advanced constraint options that provide even more control over the model-building process and outcomes

ModelCat automatically handles the complexity of turning those decisions into a working model, while still giving you control where it matters. In addition, ModelCat provides an Advanced Mode that gives experienced users deeper control over architecture and optimization decisions, allowing teams to guide the model-building process where needed.

How does ModelCat support my ML or engineering team?

ModelCat is designed to amplify the capabilities of your existing ML, engineering, and product teams — not replace them.

Traditional AI development often forces data scientists to spend large amounts of time manually tuning architectures, debugging deployment issues, and iterating through hardware constraints. ModelCat automates much of that heavy experimentation so teams can move faster from data to production-ready models.

This allows data science teams to work in lockstep with engineering and product teams, providing end-to-end AI solutions instead of handing off models that still require extensive optimization before deployment.

With ModelCat handling architecture exploration, optimization, and hardware validation, your team can focus on higher-value work such as:

Designing better datasets and training strategies
Exploring new product capabilities powered by AI
Improving system performance and reliability
Developing more advanced models and features

This also reflects a broader shift in how organizations approach AI. The real value — and competitive advantage — lives in your data. ModelCat makes model building a more mechanical and reliable process, allowing teams to double down on improving and leveraging their data rather than repeatedly rebuilding models from scratch.

In practice, ModelCat becomes a powerful assistant for your team, accelerating development while keeping humans firmly in control of system design and product direction.

What hardware does ModelCat support?

ModelCat supports a growing list of chips from a range of edge silicon vendors, with new ones being added continuously. ModelCat is inherently agnostic to the chip architectures used for inference, allowing it to support processors with NPUs, different instruction sets, and other acceleration designs.

If you don’t see your chip listed, ModelCat can onboard new hardware quickly using our AI-powered chip profiling process.

View supported chips.

Where are the NXP chips?

We have collaborated with NXP to build eIQ® Model Creator, an NXP-specific version of ModelCat. You'll find a comprehensive list of NXP silicon in eIQ Model Creator.

How does ModelCat validate models?

ModelCat validates models by executing them on real hardware and then measuring the results.

By testing on the same exact hardware that you are targeting for your project, ModelCat ensures:

Performance claims are reliable
Tradeoffs are visible early
Models behave predictably in production environments

Models are validated against real hardware constraints, including memory, power, and performance, inside ModelCat's agent loop, and the results are presented to the user in ModelCat's UI.

Learn how validation happens on ModelCat’s hardware farm.

Are there limits on the size of models ModelCat can produce?

ModelCat does not impose a fixed limit on model size. Instead, model size is determined by the constraints of the deployment environment and the goals defined by the user.

Since ModelCat builds models specifically for the target hardware, it automatically balances factors such as memory usage, compute requirements, latency, and power consumption when designing architectures.

For edge and embedded systems, this often means producing smaller, highly efficient models that fit within strict device constraints. For more powerful hardware, ModelCat can explore larger architectures if they better meet the performance objectives defined by the user.

In practice, ModelCat focuses less on maximizing raw model size and more on generating models that deliver the best possible performance for the hardware they will run on.

How is ModelCat different from AutoML tools?

AutoML tools typically focus on improving model training efficiency or maximizing benchmark accuracy. ModelCat focuses on ensuring models actually work in production environments, particularly on edge and embedded hardware.

Instead of optimizing only for training metrics, ModelCat builds models around real-world production constraints such as memory, latency, power consumption, and target hardware behavior. This allows teams to move directly from data to production-ready models without the long iteration cycles common in traditional AutoML workflows.

Can ModelCat be used alongside existing tools?

Yes. ModelCat is designed to work seamlessly with other tools. The system integrates into modern development workflows and complements existing tools rather than replacing them outright. Since ModelCat makes all artifacts available to the user, teams can take a ModelCat-generated AI model and continue training or other experimentation in other environments if desired.

ModelCat also provides output models in the exact format required by your selected hardware SDK or supporting libraries. This includes workflows within ecosystems such as NXP’s eIQ AI software environment.

ModelCat is also developing API access for key capabilities such as Model Build and Model Retargeting. These APIs will allow teams to integrate ModelCat directly into automated workflows, CI/CD pipelines, and internal tooling.

If API access is important for your workflow, contact us to learn more about upcoming capabilities and early access opportunities.

Does ModelCat provide the compute infrastructure needed to build models?

Yes. ModelCat includes the infrastructure required to train, optimize, and validate models as part of the platform.

Many teams struggle to maintain dedicated training clusters or specialized infrastructure for AI experimentation. Managing GPUs, scaling workloads, and maintaining the surrounding infrastructure often requires dedicated engineering resources.

ModelCat removes that burden by including managed compute environments designed specifically for the model build process. This allows teams to focus on defining their problem, constraints, and goals rather than managing the infrastructure required to run large-scale experiments.

Why do AI models that perform well in development often fail in production?

Many AI models are developed and optimized in environments that do not reflect the constraints of real-world deployment.

In production systems—particularly on edge or embedded hardware—models must operate within strict limits around latency, power consumption, memory, and hardware capabilities.

Traditional workflows often discover these constraints late in the process, forcing teams to redesign models or compromise performance.

ModelCat addresses this problem by incorporating real deployment constraints into the model design process from the beginning. This allows teams to build AI models that are ready for production rather than discovering limitations after months of development.

How do we get started?

The best way to get started is to discuss your use case with a ModelCat application engineer. In a short call, we can review your requirements, target hardware, datasets, and deployment goals. From there, our team can show you how ModelCat generates production-ready AI models optimized for your specific constraints and environment.

Schedule a 30-minute call with a ModelCat application engineer to start your test drive.

Take a Test Drive