# alibaba/MNN

MNN: A blazing-fast, lightweight inference engine battle-tested by Alibaba, powering high-performance on-device LLMs and Edge AI.

Repository: https://github.com/alibaba/MNN
Canonical: https://ross.abutalabs.com/products/mnn
Language: C++
License: Apache-2.0
License Family: permissive
Topics: mnn, convolution, vulkan, embedded-devices, winograd-algorithm, machine-learning, deep-learning, ml, arm, llm, transformer
Last push: 2026-08-25T07:22:34+00:00

## Health v2 (maintenance only)
Score: 93/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 99, release rhythm 82, longevity 100
- inputs: {"age_days": 2697, "days_push": 8, "days_rel": 42, "gap_med": 82.5, "n_releases_24m": 9}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 15973, forks 2417 (observed 2026-08-28T04:11:14.580591+00:00)

## What it is
MNN is a lightweight, high-performance deep learning inference engine developed by Alibaba, supporting LLMs, vision, audio, and multimodal models on mobile and embedded devices. It includes backends for CPU, GPU (Vulkan, Metal), and Qualcomm Hexagon DSPs, plus companion apps like MNN Chat for on-device LLMs.

## Use cases
- run LLMs locally on Android or iOS phones
- deploy deep learning models on embedded ARM devices
- accelerate inference with Vulkan or Hexagon DSP backends
- build an offline chatbot app with on-device ASR and TTS
- run vision models like image editing on mobile
- convert and quantize models for edge deployment

## When to choose
- you need fast on-device inference on mobile or embedded hardware
- you want to run LLMs, ASR, TTS, or vision models fully offline
- you need broad backend support including GPU and DSP acceleration
- you want a battle-tested engine with production use at Alibaba

## When to avoid
- you need server-side GPU training or fine-tuning rather than inference
- you want a Python-first training framework like PyTorch
- your target is cloud deployment with abundant compute
- you need a model training pipeline, not a runtime

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, machine-learning, deep-learning, computer-vision, speech-recognition, tts, gpu-computing
- domain: deep-learning, machine-learning, large-language-models, mobile-development, embedded-systems, computer-vision
- platform: cross-platform, cpp, embedded
- tags: on-device-inference, edge-ai, mobile-llm, vulkan, qualcomm-hexagon, transformer, android, ios, gpu

## Member repositories
- alibaba/MNN (main) score 93

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:11:14.580591+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:05:18.344125+00:00, confidence not recorded.
  - readme: https://github.com/alibaba/MNN (fetched 2026-08-28T04:11:14.580591+00:00, sha 5b8164a93b13)
- Data as of 2026-08-30T08:39:29.467469+00:00.