LLamaStudio

Settings

Settings saved on the computer running LlamaStudio. Chat generation options remain in Chat Settings.

Sandbox

Sandbox directory

Default directory for web chat and agent tools. TUI sessions can select their own directory.

Restrict file tools to this directory

When disabled, file tools may access paths outside this directory.

Allow shell commands

Enabled commands run with your account's permissions. This directory sets their starting location; file restrictions do not isolate shell commands.

Changes apply to the next request. A response in progress keeps its original directory and permissions.

{% if server_running %}{{ current_model_name }}{% else %}No model loaded{% endif %}

Welcome to LLamaStudio

Please load a model to start playing with local LLMs.

Thinking

Model Details	Arch	Size	Quant	Status

File Path

File Size

Quantization

Context Length

Max context size in tokens

2048

GPU Offload

Layers to offload (999 = Max VRAM)

999

Target Device

Target specific GPU or multi-GPU split

CPU Threads

Size of the thread pool

Parallel Slots

Concurrent request slots (1 for dedicated agent/chat)

Flash Attention

Speed up context processing

Use mmap()

Enable fast virtual memory maps

CPU Mode Only

Force running strictly on host CPU

K Cache Quant

V Cache Quant

Seed Override -1 represents completely random

ROPE Freq Base

ROPE Freq Scale

Override KV e.g. qwen35.context_length=int:262144

Vision Projector Sibling mmproj GGUF is detected automatically

Chat Template Override

Custom Jinja Template

System Prompt

Temperature 0.7

Top P 0.9

Top K 40

Min P 0.05

Repeat Penalty 1.1

Max Generation Tokens

Stop Sequences (comma-separated)

Explore Open Source LLMs

Search the Hugging Face Hub for community-quantized GGUF models. Download them directly with a single click.

Server Execution Logs

Real-time output stream from the active llama-server process.