Best local LLM for your Mac or PC
Pick your machine and see which open models fit, which quant to download, and roughly how fast they run.
Your machine
Chat and writing
- 1Qwen3.8 27BQwen · 27.8B · Apache 2.0~10 tok/sSlowQ4 · 17 GB download · Dense 27B language model with a built-in vision encoder (27.8B weights in total). 262K context natively. · Official page
- 2Qwen3.6 27BQwen · 27B · See model card~10 tok/sSlowQ4 · 16 GB download · Check the license on the model card before commercial use. · Official page
- 3Muse Glimmer 30BMeta · 29.6B · Apache 2.0~9 tok/sSlowQ4 · 18 GB download · Dense 29.6B including a 1.8B vision encoder. Meta's 4-bit builds target 24GB to 32GB of VRAM. · Official page
- 4Gemma 4 31BGoogle · 30.7B · Apache 2.0~9 tok/sSlowQ4 · 18 GB download · Official page
Coding
- 1Qwen3.8 27BQwen · 27.8B · Apache 2.0~10 tok/sSlowQ4 · 17 GB download · Dense 27B language model with a built-in vision encoder (27.8B weights in total). 262K context natively. · Official page
- 2Recursive Flint 40B A2.5BRecursive · 39.8B · Apache 2.0~62 tok/sFastQ8 · 42 GB download · Mixture of experts: 39.8B total, about 2.5B active per token, 262K context. Coding-agent model released Oct 21, 2026. Recursive's 4-bit builds are 22.6 to 26 GB and it recommends 48 GB of memory. · Official page
- 3Qwen3.6 27BQwen · 27B · See model card~10 tok/sSlowQ4 · 16 GB download · Check the license on the model card before commercial use. · Official page
- 4Muse Glimmer 30BMeta · 29.6B · Apache 2.0~9 tok/sSlowQ4 · 18 GB download · Dense 29.6B including a 1.8B vision encoder. Meta's 4-bit builds target 24GB to 32GB of VRAM. · Official page
Small and fast
- 1Gemma 4 E4BGoogle · 8B · Apache 2.0~34 tok/sFastQ8 · 8 GB download · Mixture of experts: 8B total, 4.5B active per token, so it runs fast for its size. · Official page
- 2Ministral 3 8BMistral · 8B · Apache 2.0~19 tok/sUsableQ8 · 8 GB download · Official page
- 3Qwen3 8BQwen · 8.2B · Apache 2.0~19 tok/sUsableQ8 · 9 GB download · Official page
- 4Qwen3 4BQwen · 4B · Apache 2.0~39 tok/sFastQ8 · 4 GB download · Official page
Popular setups
Know the model already? See VRAM requirements for every model.
How the estimates work
Fit: a model fits if its download size plus about 1.5 GB for context fits in usable memory. On a Mac that is about two thirds of RAM (three quarters from 32GB up), since macOS keeps some for itself. On a PC it is your GPU's VRAM minus 1 GB, or 60% of RAM when running on the CPU only.
Quant: we pick the highest quality that fits: Q8 (about 8.5 bits per weight), then Q5 (5.7), then Q4 (4.8). Q4 and Q5 keep most of the quality at a fraction of the size.
Speed (estimate): generating text is limited by memory bandwidth, because every token reads the active weights once. We use bandwidth x 0.6 / bytes per token. Mixture-of-experts models only read their active experts, so they run much faster than their total size suggests. Real speed varies with the app, context length and settings, so treat it as a rough guide.
How to run a local model
Ollama: install from ollama.com, then find the model in the Ollama library and run it with one command. LM Studio: install from lmstudio.ai, search the model name, and pick the quant shown here. On a Mac, LM Studio's MLX builds are usually fastest.
Out of plan usage? Recursive Flint, a free local coding model, vs a paid plan.
FAQ
What is the best local LLM for a Mac?
It depends on RAM. With 16GB, small models like Gemma 4 E4B or Qwen3 8B run well. With 64GB, mixture-of-experts models like gpt-oss-120b don't fit, but Gemma 4 31B and Qwen3.6 27B do, and 128GB opens up gpt-oss-120b and Mistral Small 4. Pick your Mac above for the exact list.
How much RAM do I need to run a local AI model?
About 0.6 GB per billion parameters at Q4, plus a little for context. An 8B model needs about 6 GB, a 30B model about 20 GB, and a 70B model about 44 GB.
Can I detect my specs automatically?
Partly. Browsers share your platform, CPU core count, a rough RAM hint and sometimes the GPU maker. They hide the exact chip and memory size, so the finder makes a best guess and you confirm it with a tap.
Is a local model cheaper than an API?
If you already own the hardware, local models cost only electricity. For top quality, hosted models still lead; compare them on our AI API prices page.