Local AI for Windows
Lasa Logo

Run open-source AI models on your own machine.

Lasa is a Windows desktop app that runs large language and diffusion models from Hugging Face directly on your hardware. Five inference engines. Chat, code, images, and video. Private, fast, offline-capable.

Download for Windows See the engines
Lasa Screenshot
5
Inference Engines
4
Workflows
100%
Local & Private
Hugging Face
Built-in Browser
New

Five inference engines, one app

Pick the engine that fits the model and your hardware. Lasa routes each model to the right runtime automatically — and lets you override per model.

Default

llama.cpp

GGUF · CPU + CUDA · in-process

The default engine, embedded directly in Lasa via LLamaSharp. Runs GGUF files on any Windows machine — CPU or NVIDIA GPU. No server, no setup.

In-Process

ONNX Runtime GenAI

ONNX · CPU · in-process

Microsoft's quantized ONNX format, validated end-to-end with Phi-3. Excellent CPU performance on small models. Auto-detected from genai_config.json directories.

Remote

vLLM

HF safetensors · NVIDIA · OpenAI-compat HTTP

The high-throughput PagedAttention server. Run it via Docker on Windows, point Lasa at it, and chat. Validated with Gemma 3 in Docker on RTX-class GPUs.

Remote

ExLlamaV2

EXL2 / safetensors · NVIDIA · via TabbyAPI

The fast EXL2-quantized runtime, accessed through TabbyAPI's OpenAI-compatible HTTP layer. Aggressive quantization for the most VRAM-efficient quality at a given size.

Remote

TensorRT-LLM

.engine · NVIDIA · via Triton or NIM

NVIDIA's compiled-engine runtime for maximum throughput on supported GPUs. Connect to Triton Inference Server with the OpenAI frontend, or to a NIM container.

Automatic routing

Lasa inspects each model and picks the right engine — no config files, no engine flags. The active engine appears next to the model name in the composer.

Phi-3-mini • ONNX

llama.cpp and ONNX run inside Lasa with no extra setup. vLLM, ExLlamaV2, and TensorRT-LLM connect to a server you run — locally via Docker / WSL2, or anywhere reachable on the network.

Four workflows under one roof

Lasa is more than a chat box. Each tab keeps its own model list, its own conversation history, and the right tools for the job.

Chat

General-purpose conversation with any text model. Personas, system prompts, conversation history, and voice input.

Code

Programming-focused mode with syntax highlighting and tool use. Ideal for code generation, refactoring, and debugging assistance.

Image

Diffusion image generation with Stable Diffusion, Flux, and other open models. Saved straight to your ~/Lasa/Images folder.

Video

Diffusion video generation with text-to-video models. Output goes to ~/Lasa/Videos, ready to share.

Hugging Face Native

Browse and download models, in-app

Lasa includes a built-in Hugging Face browser. Search by name, switch between GGUF and ONNX formats, see file sizes and quantization variants, and download — without leaving the app.

  • Curated recommendations for Phi, Llama, Qwen, and Gemma
  • Resumable downloads that survive flaky networks
  • Per-variant downloads for ONNX directories
  • Optional HF token for gated models
Lasa - Hugging Face Browser
microsoft/Phi-3-mini-4k-instruct ONNX · 2.3 GB
google/gemma-3-1b-it GGUF · 1.1 GB
Qwen/Qwen2.5-7B-Instruct-GGUF GGUF · 4.8 GB
meta-llama/Llama-3.2-3B Downloading… 67%

Built around your privacy

Lasa is a host. You bring the models, you own the data, you control the runtime.

Privacy-First

Conversations stay on your machine. SQLite database in %LOCALAPPDATA%\Lasa, no telemetry, no third-party tracking.

Open-Source Models

Thousands of models from Hugging Face — Qwen, Llama, Phi, Gemma, Stable Diffusion, Flux, and more. Pick what fits your task and your hardware.

Offline-Capable

Once a model is downloaded, Lasa runs without an internet connection. Get AI assistance on flights, in cafés, behind firewalls.

Personas

Save reusable system prompts as personas. Switch between a code reviewer, a writing editor, and a study tutor without retyping setup every time.

Voice Input

Built-in microphone capture for hands-free prompting. Useful for long inputs and accessibility.

CLI Included

Lasa ships with a command-line interface for scripting, automation, and listing or exporting conversations from the terminal.

From install to first response

A few minutes, no terminal commands required.

1

Install

Download Lasa-1.1.0-Setup.exe and run it. Activate with a license key or start the free trial.

2

Pick a model

Open the Hugging Face browser. Search, pick a quantization, download. Or import a folder you already have.

3

Choose an engine

Lasa picks one automatically. Want vLLM or TensorRT-LLM speed? Connect a remote endpoint in two clicks.

4

Chat

Type, attach files, dictate by voice. Switch tabs to write code, generate images, or render video.

System requirements

Lasa runs on modest hardware and scales up with whatever you have.

Minimum
  • OSWindows 10 or later (64-bit)
  • CPUx64 with AVX2
  • RAM8 GB
  • Disk~10 GB for app + small models
  • GPUNone (CPU inference)
Recommended
  • OSWindows 11
  • CPUModern x64, 8+ cores
  • RAM16-32 GB
  • DiskSSD with 50+ GB free
  • GPUNVIDIA, 8 GB+ VRAM

For vLLM, ExLlamaV2, or TensorRT-LLM, you'll also need a server reachable from the machine running Lasa — Docker Desktop with WSL2 is the easiest path on Windows.

Download Lasa

Get Lasa running on your Windows machine in seconds.

Download for Windows

Requires Windows 10 or later

Get Lasa

Purchase a license key to unlock the full power of Lasa on your Windows machine.

$25