Ollama

XDA ran the same 8B local model on an RTX 5070 laptop and a machine with integrated graphics and reported the token-per-second gap was smaller than most people would guess. That is the story of 2026 local LLMs: quantization, Vulkan backends, and Apple’s Metal have pulled a lot of workloads into a range where an iGPU or a small dGPU is enough. You do not need a $2,000 rig to run useful models locally.

This guide covers eight apps for running local LLMs on integrated graphics, tested on an Intel Core Ultra 5 with Arc iGPU, an AMD Ryzen 7 with Radeon 780M, and an Apple M2. All eight run models in the 7B to 14B range at usable speeds (10 tokens per second or better) with the right quantization. We prioritized ease of setup, model compatibility, and honest performance on iGPU-class hardware.

What to look for in a local LLM runner for iGPU

Quick comparison

App Best for Free plan Starting price UI style
Ollama CLI + API for local models Yes Free (open source) Terminal or third-party clients
LM Studio Polished desktop UI Yes Free Native desktop
GPT4All Beginner-friendly desktop Yes Free Native desktop
Jan Fully open-source ChatGPT clone Yes Free Native desktop (Electron)
Msty All-in-one with RAG Yes Free base Native desktop
KoboldCpp Roleplay and creative writing Yes Free (open source) Web UI
Text Generation WebUI Power-user tinkering Yes Free (open source) Web UI
AnythingLLM Local RAG over documents Yes Free (open source) Native desktop

The apps

1. Ollama, best CLI and API runner

Ollama is the closest thing to a standard for local LLMs. Install it, run ollama run llama3.2:3b, and you have a working model in under a minute. Under the hood it uses llama.cpp with a curated model library, an OpenAI-compatible API on localhost:11434, and per-model memory management.

Where it falls short: No built-in chat UI; you use ollama run in a terminal or pair with a client (Open WebUI, Msty, or Enchanted on Mac). Model library on ollama.com is curated but limited compared to raw Hugging Face access.

Pricing:

Platforms: Windows, macOS (Apple silicon and Intel), Linux.

Download: ollama.com/download or via Homebrew / apt.

Bottom line: The best backend, full stop. Install it even if you plan to use a different UI on top.

2. LM Studio, best polished desktop UI

LM Studio wraps llama.cpp in a native desktop app with a built-in Hugging Face model browser, one-click downloads, per-model performance benchmarks, and an OpenAI-compatible API server. GPU offload settings are exposed with sane defaults for Intel Arc, AMD, and Apple silicon.

Where it falls short: Closed-source (though free). Some users report the GUI eats more RAM than the model itself on lower-end iGPUs.

Pricing:

Platforms: Windows, macOS, Linux (beta).

Download: lmstudio.ai

Bottom line: The best starting point for iGPU users who prefer a GUI. Ships with the right defaults for consumer hardware.

3. GPT4All, best beginner-friendly desktop

GPT4All by Nomic is the most approachable of the pack. Install, pick a model from the sidebar, and it downloads and runs. LocalDocs pane adds simple RAG over a folder of files without configuration.

Where it falls short: Smaller model catalog than LM Studio. Slower on Windows than Ollama or LM Studio in our tests.

Pricing:

Platforms: Windows, macOS, Linux.

Download: nomic.ai/gpt4all

Bottom line: The pick to hand your less-technical family member. Gets them running a local model without touching a terminal.

4. Jan, best fully open-source ChatGPT clone

Jan is a fully open-source desktop app that mimics ChatGPT’s interface. It runs local models via a bundled llama.cpp, connects to remote APIs (OpenAI, Anthropic, Groq, Mistral) as an option, and stores every conversation locally.

Where it falls short: Electron-based, so RAM footprint is heavier than native apps. Multi-turn conversations occasionally lose context on iGPU due to aggressive context truncation.

Pricing:

Platforms: Windows, macOS, Linux.

Download: jan.ai

Bottom line: The most open ChatGPT-alike. Best if you want a single app for local plus optional remote.

5. Msty, best all-in-one with RAG

Msty combines local models, remote APIs, RAG over documents, and split-view conversations in one app. Its “Knowledge Stacks” feature indexes folders of PDFs, Word docs, and text files so you can chat with them without setting up a vector database.

Where it falls short: Closed-source. Free tier caps some advanced RAG features.

Pricing:

Platforms: Windows, macOS, Linux.

Download: msty.app

Bottom line: The pick if you want document chat without wiring up a separate vector store.

6. KoboldCpp, best for roleplay and creative writing

KoboldCpp started as a fork of llama.cpp aimed at long-form creative writing and roleplay. It has a web UI with world info, character memory, and lore books that other runners do not surface. Backend supports Vulkan for iGPU.

Where it falls short: UI is dense with sliders that only mean something to power users. Not really designed for chat assistant workflows.

Pricing:

Platforms: Windows, macOS, Linux.

Download: KoboldCpp GitHub releases.

Bottom line: Niche pick for writers and roleplayers who want persistent characters.

7. Text Generation WebUI, best power-user tinkering

Text Generation WebUI (oobabooga) is the tinkerer’s playground. It supports llama.cpp, ExLlama, Transformers, and more backends, exposes every generation parameter, and integrates with LoRAs and fine-tuning.

Where it falls short: Setup is not for beginners; expect a 30-minute install with Python venv wrangling. The web UI is functional but not designed.

Pricing:

Platforms: Windows, macOS, Linux via Python.

Download: oobabooga/text-generation-webui GitHub.

Bottom line: The right pick when you know what you want to tune and cannot get to it in a nicer app.

8. AnythingLLM, best local RAG over documents

AnythingLLM is a purpose-built RAG app that pairs a local (or remote) LLM with a document store. Point it at a folder or drop files in, and it indexes them into a local vector database. Multi-user workspaces separate contexts for different projects.

Where it falls short: RAG-focused; pure chat is possible but not the point. Setup involves picking an embedding model separately from the chat model, which trips up beginners.

Pricing:

Platforms: Windows, macOS, Linux, plus a Docker deployment for self-hosting.

Download: anythingllm.com or Mintplex-Labs/anything-llm GitHub.

Bottom line: The pick if the reason you want local AI is to query your own document collection privately.

How to pick the right one

Best paired setup for most iGPU users: Ollama as the backend + LM Studio or Open WebUI as the client. That gets you the fastest backend with the friendliest interface.

FAQ

What is the best local LLM app for Intel Arc iGPU?

LM Studio and Ollama both use Vulkan and work well on Intel Arc iGPUs (Xe, Xe2, Xe-LPG). Model choice matters more than app choice; Q4_K_M quantizations of 7B to 8B models are the sweet spot.

Can I run local LLMs on a Ryzen APU without a discrete GPU?

Yes. Ryzen 7000 and 8000 series Radeon 780M / 890M iGPUs run 7B models at 10 to 20 tokens per second in Q4 quantization via Vulkan. Increase system RAM allocation to graphics in BIOS if you want to load bigger contexts.

What is the best local LLM for a MacBook with M2 or M3?

Apple silicon runs Llama 3.1 8B, Qwen 3 8B, and Mistral 7B at 20 to 40 tokens per second on M2. LM Studio and Ollama both use Metal automatically. On 16GB unified memory, stick to Q4_K_M or Q5_K_M.

Do I need to know Python to run local models?

No. Ollama, LM Studio, GPT4All, Jan, and Msty are all one-click installs with no Python required. Text Generation WebUI is the exception.

Which local LLM app is fully open source?

Ollama, GPT4All, Jan, KoboldCpp, Text Generation WebUI, and AnythingLLM are all open source. LM Studio and Msty are free but closed source.

How much RAM do I need for local LLMs?

For 7B to 8B models in Q4 quantization, 16GB total system RAM is workable and 32GB is comfortable. iGPU shares system RAM, so budget accordingly. For 13B to 14B models, plan for 32GB minimum.