The models you run at home come from labs in the United States, China, Europe and beyond. Some give them away freely, some attach conditions, and the best known, like Claude and ChatGPT, don’t give them away at all.
The models in our checker
Every model our checker sizes up, with where it’s from and what it can do. Pick a tag to narrow the list.
12 models
Llama 3.2 3B
Meta
Quick answers and summaries.
- Size
- 3.21B · 2 GB download
- Downloads
- 87M for all of llama3.2
Gemma 4 E4B
Google
Small, and can see and hear.
- Size
- 8B, 4.5B active · 9.5 GB download
- Downloads
- 28M for all of gemma4
Qwen3 8B
Alibaba (Qwen)
A capable all-rounder.
- Size
- 8.19B · 5.2 GB download
- Downloads
- 42M for all of qwen3
Llama 3.1 8B
Meta
Widely supported by tools.
- Size
- 8.03B · 4.9 GB download
- Downloads
- 121M for all of llama3.1
Gemma 3 12B
Google
Writing, and reading images.
- Size
- 12.2B · 8.1 GB download
- Downloads
- 42M for all of gemma3
gpt-oss 20B
OpenAI
Reasoning and using tools.
- Size
- 20.9B, 3.6B active · 14 GB download
- Downloads
- 14M for all of gpt-oss
Mistral Small 3.2
Mistral AI
A strong general model.
- Size
- 23.6B · 15 GB download
- Downloads
- 2.5M for all of mistral-small3.2
Gemma 4 26B-A4B
Google
Big-model answers at small-model speed.
- Size
- 25.2B, 3.8B active · 19 GB download
- Downloads
- 28M for all of gemma4
Qwen 3.6 27B
Alibaba (Qwen)
Coding and hard problems.
- Size
- 27B · 19 GB download
- Downloads
- 7.4M for all of qwen3.6
Gemma 4 31B
Google
Google’s best open model.
- Size
- 30.7B · 20 GB download
- Downloads
- 28M for all of gemma4
Qwen 3.6 35B-A3B
Alibaba (Qwen)
Fast, capable, and light on memory for its size.
- Size
- 35B, 3B active · 23 GB download
- Downloads
- 7.4M for all of qwen3.6
Llama 3.3 70B
Meta
Close to the big cloud models.
- Size
- 70.6B · 43 GB download
- Downloads
- 4.4M for all of llama3.3
No model in the checker matches both. Try fewer filters.
- Published: From the model's published architecture.Sizes and capabilities: each model’s published card and configChecked: 31 Oct 2026
- Spec sheet: From the manufacturer's published specifications.Downloads: Ollama library, counted for the whole familyChecked: 31 Oct 2026
- Sees images
- You can show it photos, screenshots and documents.
- Hears audio
- It can listen to speech and sound, not just read text.
- Thinks first
- It can work through a problem step by step before answering. Slower, but better on hard questions.
- Uses tools
- Apps can let it search, run code or use other programs to get things done.
- Long documents
- It can take in 128K tokens or more at once, if your memory allows: a long report, or a folder of code.
- Many languages
- Trained to read and write dozens of languages or more, not just English.
Open, open-ish and closed
“Open” gets used loosely. What matters is the licence: the terms you agree to when you download a model.
Open source
A standard licence such as Apache 2.0 or MIT. Use it, change it, build a business on it.
For example Qwen, Gemma 4, gpt-oss, Mistral, Phi, OLMo
Open, with conditions
Free for most people, under the maker’s own terms: credit rules, limits for very large companies, or uses it forbids.
For example Llama, Kimi, Nemotron, Falcon
Research only
Free to try and study, but not for paid work without a separate deal.
For example Most FLUX models, Cohere Command, EXAONE
Closed
Only available as a cloud service. You can’t download it, so you can’t run it at home.
For example Claude (Anthropic), ChatGPT (OpenAI), Gemini (Google), Grok (xAI)
Strictly, most are Open-weight modelA model whose weights are published for anyone to download and run.Read the full entry, not open source: you get the finished model, not the data or code behind it. A few labs, like Ai2 with OLMo, publish everything. For running a model at home, the licence is what counts.
Mac, PC or both?
Almost every open model runs on both: the same file works in Ollama on a Mac or a PC. What changes is how big a model fits and how fast it writes, which the Hardware page covers.
Only on a Mac
- Apple’s own model Apple
The small model behind Apple Intelligence, built into macOS 26 and up. Apps can use it, but you can’t download it or run it anywhere else.
- MLX versions of models Apple’s format
Copies of open models converted for Apple’s MLX. Often faster on a Mac. The model inside is the same one PCs run.
- Draw Things An app
A free app that makes pictures, and short videos, on Macs, iPads and iPhones.
On both
Nearly everything else.
Every model in our checker, every model maker below, and the speech and picture models on Local vs cloud.
Video is the exception in practice: it runs on Macs, but so slowly that most people use an NVIDIA card.
Only on a PC
- Phi Silica Microsoft
A small model built into Windows on Copilot+ PCs. It runs on the chip’s AI engine, and only there.
- NVIDIA-only tools NVIDIA
Software built on CUDA, NVIDIA’s platform, such as TensorRT and many video add-ons. They need an NVIDIA card.
- Spec sheet: From the manufacturer's published specifications.Apple: Foundation Models frameworkChecked: not yet verified
- Spec sheet: From the manufacturer's published specifications.Microsoft: Phi SilicaChecked: not yet verified
- Unverified: A stand-in value we have not confirmed yet.Model makers’ own pages, linked belowChecked: not yet verified
Who makes them
The main makers of downloadable models, by where they’re based. “At home” says which of their models a home computer can run.
United States
- MetaOpen, with conditions
Llama 4, Llama 3 · United States
At home Llama 3 at 3B and 8B on almost anything; 70B on a large Mac or a big graphics card.
The model that started the local movement. Meta’s newest flagship models are cloud-only.
Their models, from Meta - GoogleOpen source
Gemma 4, Gemma 3 · United States
At home Every size, from the tiny ones for phones up to about 30B.
Gemma 4 moved to the Apache 2.0 licence; earlier Gemmas have Google’s own terms. Many can read images.
Their models, from Google - OpenAIOpen source
gpt-oss · United States
At home gpt-oss 20B on a 16 GB graphics card or a 24 GB Mac. The 120B needs a Mac with 96 GB or more.
The makers of ChatGPT. Also made Whisper, the speech-to-text model most dictation apps use.
Their models, from OpenAI - MicrosoftOpen source
Phi-4 · United States
At home Every size, on most machines: Phi models are small by design.
Also builds Phi Silica into Copilot+ PCs, where it runs on the chip’s AI engine.
Their models, from Microsoft - NVIDIAOpen, with conditions
Nemotron, Parakeet · United States
At home Nemotron Nano, about 30B, on a 24 GB card or a 32 GB Mac. Parakeet transcribes speech on almost anything.
Publishes much of its training data too. Its models run on Macs as well as on its own cards.
Their models, from NVIDIA - IBMOpen source
Granite 4 · United States
At home The small ones on almost anything; the largest, about 32B, on a 24 GB card.
Aimed at business use, with clear records of what the models were trained on.
Their models, from IBM - Ai2Open sourceData and code too
OLMo 3 · United States
At home 7B on almost anything; 32B on a 24 GB card or a 48 GB Mac.
A non-profit research institute. It publishes the training data and code as well, so anyone can see exactly how the model was made.
Their models, from Ai2 - Nous ResearchOpen, with conditions
Hermes 4 · United States
At home 14B on a 12 GB card or a 24 GB Mac; 70B on a Mac with 64 GB.
Builds on other companies’ models. The 14B is Apache 2.0, from Qwen; the larger ones follow Meta’s Llama terms.
Their models, from Nous Research
China
- Alibaba (Qwen)Open source
Qwen 3.6, Qwen 3; Wan for video · China
At home The widest range of sizes of anyone, from under 1B to about 30B, plus larger ones for big Macs.
Among the most downloaded open models in the world. Its Wan models are the leading open video makers.
Their models, from Alibaba (Qwen) - DeepSeekOpen source
DeepSeek V4, R1 · China
At home Too big for home computers. The smallest V4 still has about 280 billion parameters.
Its R1 model, in early 2025, showed an open model could rival the best cloud ones. You will mostly meet it as a cloud service.
Their models, from DeepSeek - Moonshot AIOpen, with conditions
Kimi K2, K3 · China
At home Too big for home computers: about a trillion parameters and up.
Among the strongest open models for coding and long tasks. Its licence asks very large companies to credit it.
Their models, from Moonshot AI - Z.ai (Zhipu)Open source
GLM-5, GLM-4.5 · China
At home GLM-4.5 Air, at about 100B, on a Mac with 96 GB or more. The rest are too big.
Strong at coding and at using tools.
Their models, from Z.ai (Zhipu)
Europe
- Mistral AIOpen source
Mistral Small, Ministral 3, Mistral Large 3 · France
At home Ministral at 3B, 8B and 14B, and Mistral Small 3.2 at 24B. Mistral Large is far too big.
Europe’s leading model maker. Most of its open models are Apache 2.0.
Their models, from Mistral AI - Black Forest LabsResearch only
FLUX.2, FLUX.1 · Germany
At home FLUX.2 klein 4B on a 12 GB card or a 16 GB Mac. The bigger FLUX models want 24 GB.
Makes pictures, not text. The small klein 4B is Apache 2.0; most of the rest are for non-commercial use only.
Their models, from Black Forest Labs
Canada, the Gulf and Asia
- CohereResearch only
Command A, Command R · Canada
At home Command R 7B on most machines; the larger ones need a big Mac or card.
Built for business search and documents, in many languages. The free weights are for research.
Their models, from Cohere - TIIOpen, with conditions
Falcon H1 · United Arab Emirates
At home Most sizes, which start well under 1B.
A government research centre in Abu Dhabi. Its Falcon models include ones built for Arabic.
Their models, from TII - LG AI ResearchResearch only
EXAONE 4 · South Korea
At home The 1.2B and 32B versions; the newest is far too big.
Strong in Korean and English.
Their models, from LG AI Research - Sarvam AICheck the licence
Sarvam 30B, 105B · India
At home 30B on a large Mac or a big card, if it fits; we have not checked yet.
Built for Indian languages.
Their models, from Sarvam AI
No makers match.
Licences and line-ups change often; these stay unverified until we’ve read each licence. Before paid work, read the model’s licence on the maker’s page.
Hermes, and models built on models
Training a model from scratch costs millions, so smaller labs often take an open model and Fine-tuneA model trained further, on top of someone else’s, to change how it behaves.Read the full entry it: train it further to change how it behaves.
Hermes, from Nous Research in the United States, is the best-known example. Hermes 4 comes in three sizes:
- 14B, built on Alibaba’s Qwen3 14B, under the Apache 2.0 licence.
- 70B and 405B, built on Meta’s Llama 3.1, under Meta’s Llama terms. The 405B is far too big for home.
Hermes is tuned to follow instructions closely and refuse less, with a reasoning mode you can switch on or off. People who find the big companies’ models too cautious often like it.
Hermes 4 14B at 4-bit with 8K of conversation runs well on 7 of 10 of the machines in our checker.
| Machine | Verdict | Needs | Tokens/s |
|---|---|---|---|
| Mac mini M4, 16 GB | Won't fit | 10.9 of 10.7 GB | — |
| MacBook Air M4, 24 GB | Runs well | 10.9 of 16.1 GB | 5.6–9.3 |
| MacBook Pro M4 Pro, 24 GB | Runs well | 10.9 of 16.1 GB | 13–21 |
| MacBook Pro M4 Pro, 48 GB | Runs well | 10.9 of 36 GB | 13–21 |
| MacBook Pro M4 Max, 64 GB | Runs well | 10.9 of 48 GB | 25–42 |
| RTX 4060 (8 GB) | Runs slowly | 10.9 of 7.2 GB | 5.8–9.6 |
| RTX 4070 (12 GB) | Fits, just | 10.9 of 11.2 GB | 24–39 |
| RTX 3090 (24 GB) | Runs well | 10.9 of 23.2 GB | 44–73 |
| RTX 4090 (24 GB) | Runs well | 10.9 of 23.2 GB | 47–78 |
| RTX 5090 (32 GB) | Runs well | 10.9 of 31.2 GB | 84–139 |
- Published: From the model's published architecture.NousResearch/Hermes-4-14B config.jsonChecked: 30 Oct 2026
- Estimate: Calculated from a stated rule of thumb. Real results vary.Fit and speed: the site’s estimatesChecked: not yet verified
Does it matter where a model is from?
- For privacy, no. A downloaded model is a file of numbers; it can’t send anything anywhere. In Ollama or LM Studio, a model from China is as private as one from California.
- For how it behaves, yes. A model reflects what its makers trained it on. Chinese models may avoid some political topics; American ones have guardrails of their own. Fine-tunes like Hermes change some of this.
- For the licence, yes. Licences differ more by company than by country, so check the one for the model you use.
- At work, maybe. Some companies and governments only allow models from certain countries. Ask before you install one on a work machine.