Best local LLM to run at home (Qwen 3.6, Qwen 3.8 and more)
Short answer: For most people with a recent PC or Mac: start with an open-weight Qwen model in Ollama or LM Studio. Qwen3.6 comes as a 27B dense model and a 35B-A3B mixture-of-experts model; Qwen3.8-27B is an Apache-2.0 dense model with 262K native context. Pick the size that fits your memory.
| Model | Type | Notes |
|---|---|---|
| Qwen3.6-27B | Dense, 27B | Vision and tool use, 256K context; ~16-19 GB at 4-bit (third-party estimate) |
| Qwen3.6-35B-A3B | MoE, 3B active | Fast for its size; ~22-25 GB at 4-bit (third-party estimate) |
| Qwen3.8-27B | Dense, 27B, multimodal | 262K native context, Apache 2.0 |
| Smaller 7-8B models | Dense | Run on CPU with 16 GB RAM, see offline Linux |
Which tool?
- Ollama: one command to download and chat. Steps.
- LM Studio: point-and-click app with a model browser.
- llama.cpp: build from source, best for servers and air-gapped machines.
Hardware
Memory decides what you can run. A 24 GB GPU or a Mac with lots of unified memory handles 27-35B models at 4-bit. See Mac Studio M5 for local AI.
Industrial or offline site? Start with the Industrial course.
Related questions
Is local AI free?
The software and open weights are free; check each model's license. You pay for hardware and power.
What is Qwen3.8-Max?
A very large hosted flagship from Alibaba (QwenCloud/Model Studio), not something most people run at home.
Where these facts come from
- Qwen3.6-27B on Hugging Face: official model card
- Qwen3.8-27B on Hugging Face: official model card, Apache 2.0
- Qwen blog: Qwen3.8: Qwen3.8 release notes
- Ollama library: qwen3.6: pull/run commands and download sizes
- LM Studio: Qwen3.6: run Qwen3.6 in LM Studio
- localmodel.run comparison: third-party memory estimates for Qwen3.6 quantizations
- Google Trends: US search interest and related queries, pulled Oct 29, 2026
Checked Oct 29, 2026. Plans and prices change, so confirm on the official page before you pay.
More answers: all questions →