Qwen
Qwen3.8 27B is the smartest model one 24 GB graphics card runs.
FreeSome setupOpen
Best local AI models (chat AI on your own computer)
Short answer: Qwen. Install LM Studio or Ollama and download Gemma 4 12B if you have 16 GB of memory. With a 24 GB graphics card or a 32 GB Mac, step up to Qwen3.8 27B.
Which AI chat model can I run on my computer?
Gemma
Gemma 4 12B is free for any use and runs on most computers.
FreeSome setupOpen
Gemma
Runs well on a 16 GB Mac or a 12 GB graphics card.
FreeSome setupOpen
gpt-oss
Fast reasoning on a cheaper 16 GB card, thanks to its design.
FreeSome setupOpen
Kimi K3.1 Pro weights came out on 2 November 2026, the strongest open model yet, but server-sized.
All 10 side by side.
Price, whether it runs on your own computer, and the main catch.
| # | Tool | Best for | Price | On your PC | Main downside |
|---|---|---|---|---|---|
| 01 | QwenAlibaba (Qwen team)Best overall | Writing, coding, documents, image questions | Free | Some setup | 27B needs a 24 GB graphics card or 32 GB Mac |
| 02 | GemmaGoogle DeepMind | Everyday chat on a normal laptop | Free | Some setup | Less clever than Qwen3.8 27B on hard tasks |
| 03 | gpt-ossOpenAI | Fast reasoning on 16 GB cards | Free | Some setup | Over a year old, newer Qwen and Gemma beat it |
| 04 | GLMZ.ai (Zhipu) | Coding help and tool use | Free | Some setup | Text only, cannot read images |
| 05 | NemotronNVIDIA | Long documents and agent tasks | Free | Some setup | Custom NVIDIA license, not a standard open one |
| 06 | Mistral (Ministral, Magistral, Devstral)Mistral AI | Chat in European languages | Free | Some setup | Small models not updated since December 2025 |
| 07 | KimiMoonshot AI | Companies hosting their own top AI | Free | Technical | Download is about 1.6 TB, needs a GPU server |
| 08 | DeepSeek R1 (distilled)DeepSeek | A thinking model on laptops | Free | Some setup | Old (May 2025); Gemma 4 and Qwen3.5 now beat it |
| 09 | PhiMicrosoft | Maths and logic questions | Free | Some setup | Almost two years old, newer models beat it |
| 10 | LlamaMeta | Old guides and tutorials | Free | Some setup | No new open Llama since April 2025 |
Nothing matches. Try another filter.
Best first, with the catch.
-
Qwen
Alibaba (Qwen team) · Qwen3.8 27B
The smartest model one gaming graphics card runs. Sizes 0.8B to 2.4T.
Best overallFreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, no account, free for business use.
- On your PC
- Some setup24 GB graphics card or Mac with 32 GB memory
- Top quality among models that fit one home computer
- Reads images and video, very long documents
- 27B needs a 24 GB graphics card or 32 GB Mac
- On smaller machines use Qwen3.5 9B (7 GB download) instead
-
Gemma
Google DeepMind · Gemma 4 12B
Google's open model. The 12B fits a normal 16 GB laptop.
Best freeBest on your computerFreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, no account, free for business use.
- On your PC
- Some setup12 GB graphics card or Mac with 16 GB memory
- Runs well on 16 GB Macs and 12 GB cards
- Understands images and short audio clips
- Less clever than Qwen3.8 27B on hard tasks
- Bigger 26B and 31B sizes need 24 GB of memory
-
gpt-oss
OpenAI · gpt-oss-20b
OpenAI's free download model. Fast, good at reasoning. 20B and 120B.
Best valueFreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, no account, free for business use.
- On your PC
- Some setup16 GB graphics card, or 24 GB Mac or RAM
- Very fast for its size (only 3.6B parts active)
- Choose low, medium or high thinking effort
- Over a year old, newer Qwen and Gemma beat it
- Text only, cannot read images
-
GLM
Z.ai (Zhipu) · GLM-4.7-Flash
Strong at coding and agents. Local size is GLM-4.7-Flash 30B.
FreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, no account, MIT license.
- On your PC
- Some setup24 GB graphics card, 32 GB Mac, or 32 GB RAM
- Fast, only about 3B parts active per word
- Good at code and long tasks with tools
- Text only, cannot read images
- Newer GLM-5.3 Flash (320B) is far too big for home
-
Nemotron
NVIDIA · Nemotron 3 Nano 30B-A3B
NVIDIA's fast thinking model with a huge memory for long files.
FreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, commercial use allowed under NVIDIA's license.
- On your PC
- Some setup24 GB graphics card, 32 GB Mac, or 32 GB RAM
- Reads up to 1 million tokens at once
- Fast, only about 3.5B parts active per word
- Custom NVIDIA license, not a standard open one
- Text only, six languages officially supported
-
Mistral (Ministral, Magistral, Devstral)
Mistral AI · Ministral 3 14B
European open models. Ministral 3 14B for chat, Devstral for code.
FreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, no account, free for business use.
- On your PC
- Some setup12 GB graphics card or Mac with 16 GB memory
- Fits a 12 GB card or 16 GB Mac
- Reads images, good in French, German, Spanish
- Small models not updated since December 2025
- Behind Qwen3.8 and Gemma 4 on hard questions
-
Kimi
Moonshot AI · Kimi K3.1 Pro
The strongest open model today, but server-sized, not for home PCs.
FreeOpenRuns on your PC- Price
- Free, open sourceWeights are a free download for your own servers.
- On your PC
- TechnicalGPU server with about 1.6 TB of graphics memory
- Top open model on Artificial Analysis index
- Reads images, 1 million token memory
- Download is about 1.6 TB, needs a GPU server
- Custom license, read it before business use
-
DeepSeek R1 (distilled)
DeepSeek · DeepSeek-R1-0528-Qwen3-8B
The famous thinking model, shrunk to 8B so laptops can run it.
FreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, no account, MIT license.
- On your PC
- Some setupAny laptop with 8 GB RAM; faster with a graphics card
- Runs on most laptops with 8 GB memory
- Shows its step by step thinking
- Old (May 2025); Gemma 4 and Qwen3.5 now beat it
- Real DeepSeek V4 models are far too big for home
-
Phi
Microsoft · Phi-4 14B
Microsoft's small model, good at maths and logic for its size.
FreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, no account, MIT license.
- On your PC
- Some setup12 GB graphics card or Mac with 16 GB memory
- Fits a 12 GB card or 16 GB Mac
- MIT license, free for any use
- Almost two years old, newer models beat it
- Short memory (16K tokens), weak on long files
-
Llama
Meta · Llama 3.1 8B
The best-known local model, now outdated. 8B runs on most laptops.
FreeOpenRuns on your PC- Price
- Free, open sourceFree download. Runs offline, under Meta's own Llama license.
- On your PC
- Some setupAny laptop with 8 GB RAM; faster with a graphics card
- Runs on most laptops with 8 GB memory
- Supported by every app and tutorial
- No new open Llama since April 2025
- Clearly behind Gemma 4 and Qwen at the same size
10 of these run on your own computer.
Private and offline. Check what your machine can handle.
| Tool | Setup | You need |
|---|---|---|
| Qwen | Some setup | 24 GB graphics card or Mac with 32 GB memory |
| Gemma | Some setup | 12 GB graphics card or Mac with 16 GB memory |
| gpt-oss | Some setup | 16 GB graphics card, or 24 GB Mac or RAM |
| GLM | Some setup | 24 GB graphics card, 32 GB Mac, or 32 GB RAM |
| Nemotron | Some setup | 24 GB graphics card, 32 GB Mac, or 32 GB RAM |
| Mistral (Ministral, Magistral, Devstral) | Some setup | 12 GB graphics card or Mac with 16 GB memory |
| Kimi | Technical | GPU server with about 1.6 TB of graphics memory |
| DeepSeek R1 (distilled) | Some setup | Any laptop with 8 GB RAM; faster with a graphics card |
| Phi | Some setup | 12 GB graphics card or Mac with 16 GB memory |
| Llama | Some setup | Any laptop with 8 GB RAM; faster with a graphics card |
Quick answers.
What is the best AI model to run locally?
For most people Gemma 4 12B, because it runs on a 16 GB laptop and handles images. If you have a 24 GB graphics card or a 32 GB Mac, Qwen3.8 27B is clearly smarter.
Can I run ChatGPT offline?
Not ChatGPT itself, but OpenAI's free gpt-oss-20b runs offline in LM Studio or Ollama on a computer with a 16 GB graphics card or about 24 GB of memory.
How much RAM do I need to run a local AI model?
About 8 GB for a small 4B to 8B model, 16 GB for a 12B to 14B model, and 24 to 32 GB for 27B to 30B models. Graphics card memory or a Mac makes them much faster.
Is running AI locally free?
Yes. The models on this page are free downloads and apps like LM Studio, Ollama and Jan are free. You only pay for your computer and electricity. Your chats never leave your machine.
Ollama or LM Studio, which is easier?
LM Studio is easier for beginners, a normal app with a model search and chat window. Ollama is lighter and better if you want other apps to use your model. Both are free.
Can I run DeepSeek locally?
The full DeepSeek V4 models need a server. On a laptop you can run the small DeepSeek R1 8B, but Gemma 4 and Qwen3.5 of similar size are now better.
Rankings we read: LMArena Text, Artificial Analysis, open models, Artificial Analysis, small open models. Prices from each official pricing page, checked 2026-11-05.