Deci

What is Deci?

Deci was an Israeli deep learning acceleration platform that made AI models run faster and cheaper without losing accuracy. Its AutoNAC engine, short for Automated Neural Architecture Construction, redesigned a model for a chosen hardware target to raise throughput and cut latency. The company also released the open-source SuperGradients training library, the YOLO-NAS object detection models and the DeciLM-7B language model.

Dr Yonatan Geifman, Professor Ran El-Yaniv of the Technion and Jonathan Elial founded Deci in late 2019, according to Ynet, and the firm raised about $55 million. Globes reported in April 2024 that NVIDIA planned to buy it, and Ynet and the ATS put the price at $300 million in early May 2024. Deci’s products are now discontinued. An NVIDIA representative wrote in 2025 that YOLO-NAS is no longer maintained, and SuperGradients had its last PyPI release, version 3.7.1, in April 2024. The deci.ai website still loads legacy marketing and pricing pages, but they describe a service that no longer develops.

Key Features

• AutoNAC engine: Deci’s AutoNAC engine searched for and built a new neural network architecture tuned to a specific task, dataset and hardware target. The company marketed it as AI crafted by AI, aiming for higher throughput and lower latency at similar accuracy.
• Runtime optimisation: Deci compiled and quantised trained models for deployment with TensorRT and OpenVINO in a few clicks. The SuperGradients README claimed gains of up to 10 times in throughput, latency, memory and model size, a vendor figure rather than an independent benchmark.
• Hardware benchmarking: Deci benchmarked a model’s performance on different hardware and batch sizes before deployment. Teams could compare Intel CPUs, NVIDIA GPUs and Jetson edge devices side by side, and a cloud cost calculator estimated serving costs.
• SuperGradients library: SuperGradients was Deci’s open-source PyTorch library for training classification, segmentation, detection and pose estimation models. It shipped under the Apache 2.0 licence with ready-made training recipes, pretrained checkpoints and quantisation-aware training notebooks.
• YOLO-NAS models: Deci released YOLO-NAS in May 2023 as an object detection model built with its architecture search technology. The SuperGradients README claimed a better accuracy and speed trade-off than YOLOv5 to YOLOv8, and a YOLO-NAS-Pose variant followed in November 2023.
• DeciNets catalogue: Deci’s Professional plan included the DeciNets catalogue of pre-optimised model architectures. Teams could start from a network already tuned for speed instead of running a full architecture search, which shortened projects for common vision tasks.
• DeciLM language model: DeciLM-7B was Deci’s efficient generative language model, published on Hugging Face. Globes described the company’s later focus as compressing generative AI language models so they ran more efficiently on NVIDIA servers.
• Team collaboration: Deci’s platform let co-workers share a model repository and track optimisation progress together. A model management hub, deployment tools and API access sat in every plan, including the free Community tier.

Pros

• Hardware-aware speed-ups: Deci tuned models for the exact chip they would run on, from Intel CPUs to NVIDIA Jetsons. That made it practical to deploy accurate vision models on edge devices with limited memory and power.
• Generous free tier: Deci’s Community plan was free with no credit card and included up to 5 runtime optimisations. Developers could benchmark models and test deployment tools before talking to sales about paid features.
• Open-source foundation: SuperGradients gave developers an Apache 2.0 training library with recipes and pretrained weights. Teams could reproduce Deci’s results, fine-tune YOLO-NAS on custom data and export to ONNX without buying the platform.
• Broad framework support: Deci worked with PyTorch, TensorFlow and ONNX models across cloud, edge and on-premises deployments, according to Cybernews. Teams did not need to rewrite existing models to try the optimisation tools.
• Strong research pedigree: Deci was founded by deep learning researchers from the Technion and counted Intel, AWS, HPE and NVIDIA as partners, according to Ynet. NVIDIA’s decision to acquire the company points to the value of its technology.

Cons

• No longer maintained: Deci’s products stopped receiving development after NVIDIA acquired the company in 2024. An NVIDIA representative said YOLO-NAS is no longer maintained, so bugs, security fixes and new hardware support will not arrive.
• Restricted YOLO-NAS weights: YOLO-NAS pretrained weights were covered by a Deci licence for research and non-commercial use only, according to the NVIDIA developer forum. Commercial users had to train from scratch or seek terms, and guidance on the forum conflicted.
• Opaque paid pricing: Deci never published prices for its Professional and Enterprise plans. Custom NAS optimisation and on-premises deployment required a sales conversation, which Cybernews flagged as a barrier for startups.
• Steep learning curve: Deci’s optimisation tools assumed solid deep learning knowledge, according to Cybernews. Small teams and solo developers could find quantisation, compilation and architecture search settings overwhelming without prior experience.
• Stale public pages: Deci’s website still shows its old pricing and product pages with no shutdown notice. New visitors may assume the service is active and only discover the status after trying to sign up or seek support.

Who is Using Deci?

• Computer vision teams: Fast detection and segmentation on edge devices.
• Enterprise AI groups: Lower inference costs in automotive and manufacturing.
• ML researchers: SuperGradients recipes and YOLO-NAS baselines.
• Edge AI developers: Models tuned for Jetson and CPU targets.

Pricing

Deci no longer sells any plan; its legacy pricing page still lists three tiers, but the products are unmaintained after the NVIDIA deal (deci.ai pricing page, checked October 2026).

Reports of the acquisition price differ, with Ynet and the ATS citing $300 million and other trackers showing other figures. Any old quote or plan detail you find online is historical only, so budget instead for a maintained optimisation tool.

What Makes Deci Unique?

Neural architecture search sat at the centre of Deci’s offer, whereas most optimisation tools only compress an existing model. AutoNAC generated a new architecture for your hardware, then the platform compiled and quantised it for deployment. Pairing that commercial engine with the open-source SuperGradients library and YOLO-NAS gave developers a free route in, before NVIDIA absorbed the team in 2024.

Compatibilities and Integrations

• Frameworks: PyTorch, TensorFlow and ONNX models.
• Deployment runtimes: TensorRT and OpenVINO compilation and quantisation.
• Hardware: Intel CPUs, NVIDIA GPUs and Jetson edge devices.
• Open source: SuperGradients on PyPI, models on Hugging Face and Roboflow.

Deci Tutorials

Deci’s tools are frozen, but archived assets can still be studied with care.

1. Install the last SuperGradients release with pip install super-gradients.
2. Load a pretrained YOLO-NAS model and run the predict function on a test image.
3. Read the YOLO-NAS licence before any commercial use of the weights.
4. Export the model to ONNX for TensorRT or OpenVINO deployment.
5. Plan a move to a maintained optimisation toolkit for production work.

How We Rated It

Summary

Deci suited machine learning teams that needed fast, accurate models on specific hardware, especially at the edge. Its main limitation today is simple: NVIDIA acquired the company in 2024 and the products are no longer maintained. SuperGradients code remains downloadable, but it receives no fixes. Check the YOLO-NAS licence first, and move production workloads to an actively supported optimisation toolkit.