Device-Native Foundation Models logopeterindia.net
Knowledge Hub

Device-Native Foundation Models

A curated index of foundation models built to run on the device itself — Edge0 language and voice models, device solutions, other on-device model families and the runtimes that run them. Offline, private and with no per-use token cost.

26 topics•5 categories•Updated November 13, 2026

What Are Device-Native Foundation Models?

3 entries

Device-native foundation models are pretrained models designed to run inference on the device itself, such as a phone, laptop, earbuds, glasses or a toy, instead of calling a cloud model. They trade parameter count and memory for low power, offline use, on-device data processing and no per-use token cost. Edge0, whose site is the basis of this page, describes its approach as on-device AI built for consumer hardware.

01Edge0On-device AI company building fast, compute-efficient foundation models for consumer hardware, with language and voice models. 02Edge0 DocsDocumentation for Edge0 models and SDKs. 03Edge0 ResearchResearch from Edge0 on on-device models.

Edge0 Models

7 entries

The Edge0 language models and Audio8 speech models, with figures as published by Edge0.

04Edge0-35B35-billion-parameter flagship sparse mixture-of-experts model that Edge0 says runs with 1-3 GB of memory by streaming expert weights from SSD. 05Edge0-8B8-billion-parameter lightweight model that Edge0 says runs with 1-2 GB of memory, for Q&A and low-to-medium complexity tasks. 06Audio8-ASR-3BSpeech recognition model that Edge0 reports at 5.04% word error rate on the open-source ASR leaderboard. 07Audio8-ASR-0.6BMid-size Audio8 speech recognition model for smaller devices. 08Audio8-ASR-0.1BSmallest Audio8 speech recognition model, for the most constrained hardware. 09Audio8-TTS-0.6B0.6B text-to-speech model that Edge0 says delivers 8B-grade voice with zero-shot voice cloning. 10Audio8-TTS-0.1BSmallest Audio8 text-to-speech model for low-power devices.

Edge0 Device Solutions

5 entries

Edge-cloud SDKs that embed Edge0 models in products.

11Smart EarbudsEdge-cloud voice SDK embedded in earbuds for natural commands and live readouts without hardware changes. 12Robotics ChipEdge-cloud human-robot interaction SDK for humanoid robots, with dialogue and command execution that work offline or on weak networks. 13Smart GlassesEdge-cloud coordination that runs frequent commands on the device and sends complex tasks to the cloud, cutting latency and cost. 14Embodied AIRobot chips plus edge voice models for offline, low-cost voice interaction aimed at makers. 15Smart ToysEdge-side models on a hybrid edge-cloud base that answer fast commands and recognise emotion.

Other On-Device Foundation Models

6 entries

Model families from other vendors that are built or sized for devices.

16GemmaGoogle's family of lightweight open-weight models, including small variants intended for phones and laptops. 17Gemini NanoGoogle's on-device Gemini model for Android, accessed through system APIs. 18Apple Foundation ModelsApple framework that gives apps access to the on-device language model behind Apple Intelligence. 19Liquid AI LFM2Liquid AI's LFM2 family of efficient foundation models designed for on-device deployment. 20MiniCPMOpenBMB's efficient language and multimodal models built for end-side devices. 21Microsoft PhiMicrosoft's family of small language models, sized to run with limited compute.

Runtimes for Running Models On-Device

5 entries

Engines that load and run models on phones, laptops and embedded hardware.

22ExecuTorchPyTorch's on-device inference runtime for mobile, embedded and edge devices. 23LiteRTGoogle's on-device runtime, formerly TensorFlow Lite. 24llama.cppC/C++ engine for running quantised language models locally on a wide range of hardware. 25MLXApple's array framework for machine learning on Apple silicon. 26MLC LLMCompiler-based engine for deploying language models natively on phones, browsers and laptops.

Related Pages

5 entries

Related research directories on this site.

27Edge AIEdge AI hardware, runtimes and on-device models. 28Small Language Models (SLMs)Compact language models for constrained environments. 29On-Device Inference EnginesEngines for running inference on devices. 30Edge-native Inference EnginesInference engines built for edge deployments. 31Edge ComputingEdge computing hub.

Frequently Asked Questions

What is a device-native foundation model?

A pretrained model designed to run inference directly on a device such as a phone, laptop, earbuds or glasses, rather than in the cloud. It works offline, keeps data on the device and avoids per-use token costs, at the price of fitting within the device's memory and power budget.

Is parameter count the same as memory use?

No. Edge0 makes this point on its site: it says its 35-billion-parameter sparse mixture-of-experts model runs with 1-3 GB of memory by streaming expert weights from SSD, while dense models of similar size need far more RAM. Memory depends on architecture, quantisation and how weights are loaded.

How do device-native models differ from small language models?

Small language models are defined by size. Device-native models are defined by where they run and are engineered around device limits, and they can include speech, vision and larger sparse models as well as small dense ones.

What is edge-cloud coordination?

It splits work between device and cloud. Frequent, simple commands run on the device for low latency and no cost, while complex tasks go to a larger cloud model. Edge0 describes this approach for earbuds, glasses and robots.

Which runtimes can run these models?

Common choices are ExecuTorch, LiteRT, llama.cpp, MLX and MLC LLM, plus vendor SDKs for specific chips. Check each model's documentation for supported formats, because availability differs by model and device.