Run Models Locally
Run open-weight models on your own computer for private data, offline work, free experiments or a local-first product.
Trends
Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.
The real choice
App or engine. Desktop apps and CLIs download models for you and expose a local API in minutes. The underlying engine gives you full control over builds, quantization and flags, and you handle the details yourself.
Pick by situation
Tap the ones that are you — the tools that fit light up below.- You want a local model behind an API for your code and dev tools in minutes
Ollama - You'd rather browse and compare models in a desktop app before writing code
LM Studio
Jan - You ship local inference inside your own app or need settings the apps don't expose
llama.cpp
- You work on a Mac with Apple silicon and want the fastest local inference or fine-tuning there
MLX LM
- You need a model family with a permissive license for commercial use
Qwen
gpt-oss
Gemma
Compare the tools
Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.
Ollamaapp · open sourcePulling and running models with one command, with a local API that other tools already support.


39+109 open source665.3k/wk−35% vs npm
Use it whenYou want a local model available to your code and dev tools with minimal setup.
Trade-offFewer tuning knobs than running the engine directly.
Used by Free AI Video Editor OpenCutAI, Gotcha, Kuky, Lagrange by OrangeCat, Morphik and 34 more · in 109 open-source projects.
Octrafic supports Ollama for fully local, private AI-powered testing. Having a simple way to run LLMs locally was key for users who can't send API specs to the cloud.
LM Studioapp · self-hostableBrowsing, downloading and chatting with models in a GUI, then serving them through a local API.


4
Use it whenYou want to compare models by hand before writing code.
Trade-offFree but not open source.
Used by Gotcha, ModelHub, Sayonic, VoiSistant.
Integrated for local, private LLM processing. LM Studio helps VoiSistant users leverage advanced on-device AI features with zero data compromise or cloud dependency.
Janapp · open sourceAn open-source desktop app for downloading and chatting with local models, with a local OpenAI-compatible server.No evidence yet★ 45k
Use it whenYou want an LM Studio-style app whose source you can read and modify.
Trade-offFewer model-management and tuning options than LM Studio.
No maker's product we track shows it for this yet.
llama.cpplibrary · open sourceFull control over builds, quantization and hardware flags, or embedding inference inside your own app.
1+27 open source339k/wk23× vs npm*
Use it whenYou're shipping local inference in a product, or need a setting the apps don't expose.
Trade-offYou compile it, convert models and pick settings yourself.
Used by LM Studio · in 27 open-source projects.
MLX LMlibrary · open sourceRunning and fine-tuning models on Apple silicon Macs, using unified memory efficiently.22 open source124.3k/wk
Use it whenYou develop on a Mac and want the most speed from it, or to fine-tune a small model locally.
Trade-offApple silicon only, and models need MLX-format weights.
No maker's product we track shows it for this yet · in 22 open-source projects.
Llamalibrary · self-hostableMeta's widely supported open-weight models, available in every local runtime and in many sizes.


21
Use it whenYou want a model family that Ollama, LM Studio and llama.cpp all run out of the box, with many community fine-tunes.
Trade-offCustom license with use restrictions rather than an open-source license; no new Llama release since Llama 4 in April 2025, as Meta's developer platform now centres on its hosted Muse models.
Used by Antispace, Bebop.ai, Integral, MindPal, NeuralAgent and 16 more.
We use Llama for a couple of our AI agents due to its open-source flexibility and efficiency. It allows us to customize the model to meet specific requirements while maintaining excellent performance.
Gemmalibrary · open sourceGoogle's small and mid-size open models, built to run on laptops and phones, with Gemma 4 under Apache 2.0.No evidence yet
Use it whenYou want a small model that runs on-device, including on mobile.
Trade-offEarlier Gemma versions carry Google's own terms of use, so check which version you ship.
No maker's product we track shows it for this yet.
gpt-osslibrary · open sourceOpenAI's open-weight reasoning models under Apache 2.0, with the 20B version running on a well-equipped laptop.No evidence yet★ 20.5k
Use it whenYou want a reasoning model that runs locally and behaves like OpenAI's API models.
Trade-offOnly two sizes, both text-only, and the 120B version needs a large GPU.
No maker's product we track shows it for this yet.
Qwenlibrary · open sourceA broad open-weight family — chat, coding, vision and embeddings — in many sizes, mostly under Apache 2.0.No evidence yet★ 27.7k
Use it whenYou want a permissive license and a size that fits your hardware, from a laptop up to a server.
Trade-offDeveloped by Alibaba in China, which matters to some customers' procurement rules.
No maker's product we track shows it for this yet.
Before you choose
Start with a small quantized model (a few billion parameters) and check speed on your actual machine before planning a feature around it. Point your code at the local OpenAI-compatible endpoint so switching to a hosted model later is a config change. Expect a big gap in quality and speed between a laptop and hosted frontier models; test on your real task.
- Choosing a model by its benchmark scores and only then finding it runs at two tokens a second on your machine.
- Assuming "open weights" means open source; some model licenses restrict commercial use, user counts or what you can build.
- Planning a local-first feature around your own high-end laptop when many users have far less memory.
Decided alongside
What the 59 makers' products here chose for their other decisions.