The models in the library
Open Models in Laghu to see them all, with each one’s size and licence. The library is grouped by family:
| Family | Made by | Sizes | Good for |
|---|---|---|---|
| Qwen2.5 | Qwen team (Alibaba) | 1.5B, 3B, 7B, 32B, 72B | Everyday chat and writing, in many languages |
| Qwen2.5 Coder | Qwen team (Alibaba) | 7B, 14B, 32B | Code, and agent mode |
| Qwen3 | Qwen team (Alibaba) | 4B, 8B, 14B, 32B | Reasoning and chat |
| Llama 3 | Meta | 1B, 3B, 8B, 70B, 405B | General chat and English writing |
| Mistral | Mistral AI | 7B, Nemo 12B, Small 24B, Large 123B | Writing, and European languages |
| DeepSeek R1 Distill | DeepSeek | 7B (Qwen), 8B (Llama) | Step-by-step reasoning and maths |
| gpt‑oss | OpenAI | 20B | Hard questions, maths, logic and planning |
| QwQ | Qwen team (Alibaba) | 32B | Long, careful reasoning |
| Small models | Hugging Face, TinyLlama | SmolLM2 1.7B, TinyLlama 1.1B | Very modest PCs |
On a laptop with 8 GB of memory, start with Qwen2.5 1.5B or Llama 3.2 3B for quick chat; with 16 GB, Qwen2.5 7B or Llama 3.1 8B. Laghu suggests one for your PC the first time it opens. How to run an LLM locally on Windows covers choosing in more detail.
Models that think before they answer
gpt‑oss 20B, the DeepSeek R1 Distill models, QwQ 32B and Qwen3 work through a problem before giving their answer. They take longer, and they are better at maths, logic, planning and tricky questions. You can read their thinking as it happens, folded away under the answer.
Models that see
Qwen2.5-VL, SmolVLM2 and the LLaVA models see pictures: attach a photo, a screenshot or a scanned page and ask about it. Qwen2.5-VL reads a picture at its own shape, which makes it the best at screenshots and pages, and the one to use for Computer use. SmolVLM2 is small enough for almost any laptop. With any model, Laghu also reads the text in an attached picture, so even a model that cannot see gets the words.
Models for agent mode
Models badged Agent in the library can use tools: read and write files, run commands and make documents in agent mode. Qwen2.5 Coder 7B is the one we recommend for it; Qwen2.5 7B, Qwen3 8B, Llama 3.1 8B and Mistral Nemo 12B work well too.
Licences
Each model is made by its own maker and comes with its own licence. Many, such as most Qwen, Mistral and gpt‑oss models, use Apache 2.0; Llama models use Meta’s community licence; a few are for non-commercial use only. Laghu shows a model’s licence before you download it, so check it if you plan to use the model for work.
Models you already have
If you have downloaded models with Ollama, Laghu finds them and marks them in the library, so nothing is downloaded twice. It also opens any .gguf model file you point it to.
Questions
Can I run DeepSeek R1 on my laptop?
Yes, the distilled versions: DeepSeek R1 Distill 7B and 8B run on an everyday laptop, offline. The full DeepSeek R1 model is far too large for a home computer and is not in the library.
Can I run Llama 70B or other very large models?
They are in the library, and Laghu runs models bigger than your free memory. They need plenty of disk space, about 40 GB for a 70B model, and on a small laptop they answer very slowly. A PC with more memory makes them practical.
Which model is best for coding?
Qwen2.5 Coder: 7B on most laptops, 14B or 32B if your PC has the memory.
Do these models work without the internet?
Yes. Once a model is downloaded, it runs entirely on your PC with the internet switched off.
Are the models free?
Yes, all of them are free to download and run. Their licences differ on commercial use, and Laghu shows each one before you download.