nexa-sdk-demo.mp4

Nexa SDK - Local On-Device Inference Framework

On-Device Model Hub | Documentation | Discord | Blogs | X (Twitter)

Nexa SDK is a local on-device inference framework for ONNX and GGML models, supporting text generation, image generation, vision-language models (VLM), audio-language models, speech-to-text (ASR), and text-to-speech (TTS) capabilities. Installable via Python Package or Executable Installer.

Features

• Device Support: CPU, GPU (CUDA, Metal, ROCm), iOS
• Server: OpenAI-compatible API, JSON schema for function calling and streaming support
• Local UI: Streamlit for interactive model deployment and testing

Latest News 🔥

• Support Nexa AI's own vision language model (0.9B parameters): nexa run omnivision and audio language model (2.9B): nexa run omniaudio
• Support audio language model: nexa run qwen2audio, we are the first open-source toolkit to support audio language model with GGML tensor library.
• Support embedding model: nexa embed <model_path> <prompt>
• Support pull and run supported Computer Vision models in GGUF format from HuggingFace: nexa run -hf <model_id> -mt COMPUTER_VISION
• Support pull and run NLP models in GGUF format from HuggingFace: nexa run -hf <model_id> -mt NLP

Welcome to submit your requests through issues, we ship weekly.

Install Option 1: Executable Installer

macOS Installer

Windows Installer

Linux Installer

curl -fsSL https://public-storage.nexa4ai.com/install.sh | sh

Currently, executable installer does not support local UI. For local UI, please install using python package below.

FAQ: cannot use executable with nexaai python package already installed

Try using nexa-exe instead:

nexa-exe <command>

Install Option 2: Python Package

We have released pre-built wheels for various Python versions, platforms, and backends for convenient installation on our index page.

CPU

pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/cpu --extra-index-url https://pypi.org/simple --no-cache-dir

Apple GPU (Metal)

For the GPU version supporting Metal (macOS):

CMAKE_ARGS="-DGGML_METAL=ON -DSD_METAL=ON" pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/metal --extra-index-url https://pypi.org/simple --no-cache-dir

FAQ: cannot use Metal/GPU on M1

Try the following command:

wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-MacOSX-arm64.sh
bash Miniforge3-MacOSX-arm64.sh
conda create -n nexasdk python=3.10
conda activate nexasdk
CMAKE_ARGS="-DGGML_METAL=ON -DSD_METAL=ON" pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/metal --extra-index-url https://pypi.org/simple --no-cache-dir

Nvidia GPU (CUDA)

To install with CUDA support, make sure you have CUDA Toolkit 12.0 or later installed.

For Linux:

CMAKE_ARGS="-DGGML_CUDA=ON -DSD_CUBLAS=ON" pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/cu124 --extra-index-url https://pypi.org/simple --no-cache-dir

For Windows PowerShell:

$env:CMAKE_ARGS="-DGGML_CUDA=ON -DSD_CUBLAS=ON"; pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/cu124 --extra-index-url https://pypi.org/simple --no-cache-dir

For Windows Command Prompt:

set CMAKE_ARGS="-DGGML_CUDA=ON -DSD_CUBLAS=ON" & pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/cu124 --extra-index-url https://pypi.org/simple --no-cache-dir

For Windows Git Bash:

CMAKE_ARGS="-DGGML_CUDA=ON -DSD_CUBLAS=ON" pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/cu124 --extra-index-url https://pypi.org/simple --no-cache-dir

FAQ: Building Issues for llava

If you encounter the following issue while building:

try the following command:

CMAKE_ARGS="-DCMAKE_CXX_FLAGS=-fopenmp" pip install nexaai

AMD GPU (ROCm)

To install with ROCm support, make sure you have ROCm 6.2.1 or later installed.

For Linux:

CMAKE_ARGS="-DGGML_HIPBLAS=on" pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/rocm621 --extra-index-url https://pypi.org/simple --no-cache-dir

GPU (Vulkan)

To install with Vulkan support, make sure you have Vulkan SDK 1.3.261.1 or later installed.

For Windows PowerShell:

$env:CMAKE_ARGS="-DGGML_VULKAN=on"; pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/vulkan --extra-index-url https://pypi.org/simple --no-cache-dir

For Windows Command Prompt:

set CMAKE_ARGS="-DGGML_VULKAN=on" & pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/vulkan --extra-index-url https://pypi.org/simple --no-cache-dir

For Windows Git Bash:

CMAKE_ARGS="-DGGML_VULKAN=on" pip install nexaai --prefer-binary --index-url https://nexaai.github.io/nexa-sdk/whl/vulkan --extra-index-url https://pypi.org/simple --no-cache-dir

Local Build

How to clone this repo

git clone --recursive https://github.com/NexaAI/nexa-sdk

If you forget to use --recursive, you can use below command to add submodule

git submodule update --init --recursive

Then you can build and install the package

pip install -e .

Differentiation

Below is our differentiation from other similar tools:

Feature	Nexa SDK	ollama	Optimum	LM Studio
GGML Support	✅	✅	❌	✅
ONNX Support	✅	❌	✅	❌
Text Generation	✅	✅	✅	✅
Image Generation	✅	❌	❌	❌
Vision-Language Models	✅	✅	✅	✅
Text-to-Speech	✅	❌	✅	❌
Server Capability	✅	✅	✅	✅
User Interface	✅	❌	❌	✅

Supported Models & Model Hub

Our on-device model hub offers all types of quantized models (text, image, audio, multimodal) with filters for RAM, file size, Tasks, etc. to help you easily explore models with UI. Explore on-device models at On-device Model Hub

Supported model examples (full list at Model Hub):

Model	Type	Format	Command
omniaudio	AudioLM	GGUF	nexa run omniaudio
qwen2audio	AudioLM	GGUF	nexa run qwen2audio
octopus-v2	Function Call	GGUF	nexa run octopus-v2
octo-net	Text	GGUF	nexa run octo-net
omnivision	Multimodal	GGUF	nexa run omnivision
nanollava	Multimodal	GGUF	nexa run nanollava
llava-phi3	Multimodal	GGUF	nexa run llava-phi3
llava-llama3	Multimodal	GGUF	nexa run llava-llama3
llava1.6-mistral	Multimodal	GGUF	nexa run llava1.6-mistral
llava1.6-vicuna	Multimodal	GGUF	nexa run llava1.6-vicuna
llama3.2	Text	GGUF	nexa run llama3.2
llama3-uncensored	Text	GGUF	nexa run llama3-uncensored
gemma2	Text	GGUF	nexa run gemma2
qwen2.5	Text	GGUF	nexa run qwen2.5
mathqwen	Text	GGUF	nexa run mathqwen
codeqwen	Text	GGUF	nexa run codeqwen
mistral	Text	GGUF/ONNX	nexa run mistral
deepseek-coder	Text	GGUF	nexa run deepseek-coder
phi3.5	Text	GGUF	nexa run phi3.5
openelm	Text	GGUF	nexa run openelm
stable-diffusion-v2-1	Image Generation	GGUF	nexa run sd2-1
stable-diffusion-3-medium	Image Generation	GGUF	nexa run sd3
FLUX.1-schnell	Image Generation	GGUF	nexa run flux
lcm-dreamshaper	Image Generation	GGUF/ONNX	nexa run lcm-dreamshaper
whisper-large-v3-turbo	Speech-to-Text	BIN	nexa run faster-whisper-large-turbo
whisper-tiny.en	Speech-to-Text	ONNX	nexa run whisper-tiny.en
mxbai-embed-large-v1	Embedding	GGUF	nexa embed mxbai
nomic-embed-text-v1.5	Embedding	GGUF	nexa embed nomic
all-MiniLM-L12-v2	Embedding	GGUF	nexa embed all-MiniLM-L12-v2:fp16
bark-small	Text-to-Speech	GGUF	nexa run bark-small:fp16

Run Models from 🤗 HuggingFace

You can pull, convert (to .gguf), quantize and run llama.cpp supported text generation models from HF with Nexa SDK.

Run .gguf File

Use nexa run -hf <hf-model-id> to run models with provided .gguf files:

nexa run -hf Qwen/Qwen2.5-Coder-7B-Instruct-GGUF

Note: You will be prompted to select a single .gguf file. If your desired quantization version has multiple split files (like fp16-00001-of-00004), please use Nexa's conversion tool (see below) to convert and quantize the model locally.

Convert .safetensors Files

Install Nexa Python package, and install Nexa conversion tool with pip install "nexaai[convert]", then convert models with nexa convert <hf-model-id>:

nexa convert HuggingFaceTB/SmolLM2-135M-Instruct

Note: Check our leaderboard for performance benchmarks of different quantized versions of mainstream language models and HuggingFace docs to learn about quantization options.

📋 You can view downloaded and converted models with nexa list

Documentation

Note

1. If you want to use ONNX model, just replace pip install nexaai with pip install "nexaai[onnx]" in provided commands.
2. If you want to run benchmark evaluation, just replace pip install nexaai with pip install "nexaai[eval]" in provided commands.
3. If you want to convert and quantize huggingface models to GGUF models, just replace pip install nexaai with pip install "nexaai[convert]" in provided commands.
4. For Chinese developers, we recommend you to use Tsinghua Open Source Mirror as extra index url, just replace --extra-index-url https://pypi.org/simple with --extra-index-url https://pypi.tuna.tsinghua.edu.cn/simple in provided commands.

CLI Reference

Here's a brief overview of the main CLI commands:

• nexa run: Run inference for various tasks using GGUF models.
• nexa onnx: Run inference for various tasks using ONNX models.
• nexa convert: Convert and quantize huggingface models to GGUF models.
• nexa server: Run the Nexa AI Text Generation Service.
• nexa eval: Run the Nexa AI Evaluation Tasks.
• nexa pull: Pull a model from official or hub.
• nexa remove: Remove a model from local machine.
• nexa clean: Clean up all model files.
• nexa list: List all models in the local machine.
• nexa login: Login to Nexa API.
• nexa whoami: Show current user information.
• nexa logout: Logout from Nexa API.

For detailed information on CLI commands and usage, please refer to the CLI Reference document.

Start Local Server

To start a local server using models on your local computer, you can use the nexa server command. For detailed information on server setup, API endpoints, and usage examples, please refer to the Server Reference document.

Swift Package

Swift SDK: Provides a Swifty API, allowing Swift developers to easily integrate and use llama.cpp models in their projects.

More Docs

Acknowledgements

We would like to thank the following projects:

• llama.cpp
• stable-diffusion.cpp
• bark.cpp
• optimum