- Category: Model Serving & Inference Engines
- Type: Edge Runtime
- License: Open source
- Region: US
- Tier: B
- First Triaged: 2026-06-16
- Last Updated: 2026-06-16
Meta; 1.0 GA Oct 2025; 50KB footprint; 12+ hardware backends; billions of users
ExecuTorch is a edge runtime in the model serving and inference engines category.
Supported model formats: pte, pytorch
Supported quantization: fp16, bf16, int8, int4
Hardware target: Edge
⏳ Benchmark results pending. See TESTING.md for methodology.
Raw results JSON:
- Official site: ⏳
- GitHub: ⏳
- Documentation: ⏳
| Date | Event | Notes |
|---|---|---|
| 2026-06-16 | First triaged | Added to roster, stub page created |
ExecuTorch is PyTorch's official edge AI runtime, enabling deployment of PyTorch models on mobile phones, embedded devices, and microcontrollers. It achieved 1.0 GA in October 2025 and supports 12+ hardware backends including Apple Silicon, Qualcomm, MediaTek, and Arm. The runtime has a ~50KB footprint and powers applications used by billions of users across Meta's apps.
ExecuTorch has 1,341 open issues — the highest of any tool in this category — reflecting its ambitious scope and rapid development. Model export from PyTorch requires careful handling of operator coverage and backend-specific delegation. The API is still evolving post-1.0, meaning breaking changes can occur.
LLM support on-device is still maturing — large model inference on phones is constrained by memory and compute. The 12+ backend support means each backend has varying levels of operator coverage and optimization. Documentation for advanced use cases (custom operators, custom backends) is still being built out.
No published security advisories. As a PyTorch Foundation project backed by Meta, it undergoes internal security review. Edge deployment introduces unique security considerations — models run on untrusted devices, requiring careful handling of model IP protection.
Rating: 6/10
ExecuTorch is the only framework that provides a unified PyTorch-to-edge pipeline across mobile, embedded, and microcontroller targets. Its delegate system allows hardware-specific acceleration (CoreML, QNN, etc.) while maintaining a single export path from PyTorch training code.
ExecuTorch is the only framework that provides a unified PyTorch-to-edge pipeline across mobile, embedded, and microcontroller targets. Its delegate system allows hardware-specific acceleration (CoreML, QNN, etc.) while maintaining a single export path from PyTorch training code.
| Rank | Tool | Stars | Key Advantage |
|---|---|---|---|
| 1 | vLLM | 85K+ | Largest community, broadest hardware support |
| 2 | SGLang | 30K | RadixAttention, best for RAG workloads |
| 3 | TensorRT-LLM | 14K | Highest single-GPU throughput on NVIDIA |
| 4 | llama.cpp | 119K | Best for CPU/consumer hardware |
| 5 | Ollama | 175K | Easiest local deployment |
See tools/README.md for the full ranking table.
Development is extremely active (pushed July 2026) with 1,063 forks. The project needs to reduce the issue backlog (1,341 open), stabilize APIs post-1.0, improve documentation for custom backends, and expand LLM support for mobile devices. Community contributions are strong given the fork count.
Maintained by the PyTorch team at Meta. Contact via GitHub Issues at pytorch/executorch, PyTorch Discussion forums (discuss.pytorch.org), or the PyTorch Slack. Core maintainers include Meta engineers working on PyTorch mobile/edge.
Best for deploying PyTorch models to mobile or edge devices, especially if you're already in the PyTorch ecosystem. For LLMs specifically, consider alongside MLC-LLM or MLX (Apple Silicon). Start with the tutorial models and verify operator coverage for your target backend.
Content for this page is licensed CC BY 4.0 — share and adapt with attribution to ArdurAI / Model Serving & Inference Engines Almanac.