Hugging Face Hub
The largest open model library. Here you find safetensors, GGUF and many other formats, with model cards and license files.
Norwegian reference for local AI — registers, plans and honest tests.
Sources for local models
Hubs, quantization sources, runtimes and Norwegian models — plus a rough memory filter and a GPU overview.
Last checked:
License first. Each model has its own license. Read the file on the model card before commercial use. See also the license register. Memory figures are approximate and depend on quantization and context length.
Modelbiblioteker og hubber
Ollama Library
Catalog of models you can fetch with ollama pull. Simple for local runs; typically GGUF under the hood.
LM Studio
Desktop app with a GUI to search, download and run local models (incl. GGUF and MLX), with a local server. A GUI alternative to Ollama.
ModelScope
Alibaba’s model hub with many Chinese and international models. A useful mirror when Hugging Face is slow.
Kaggle Models
Model catalog at Kaggle. Often needs a Kaggle account to download.
NVIDIA NGC / build.nvidia.com
NVIDIA’s catalogs for containers, SDKs and models. Relevant for CUDA and Jetson environments.
NVIDIA Jetson AI Lab
Recipes and containers for local AI on Jetson (Orin and others). Suited to edge and low power use.
Civitai
Hub for image models (Stable Diffusion etc.). Lots of user-generated content — check license and content moderation yourself.
GPT4All
App and model catalog for local chat on a regular PC. Open source (MIT for the client); models have their own licenses.
Jan
Open desktop app for local AI chat with support for several runtimes and a model catalog.
Kvantiserte varianter (Hugging Face)
Disse organisasjonene publiserer ofte ferdige GGUF/MLX-varianter. Sjekk alltid hvilken base-model og lisens checkpointen arver.
bartowski
Active Hugging Face organisation with many GGUF quantizations (often Q4/Q5/Q8) ready for llama.cpp and Ollama-like tools.
unsloth
Organisation known for efficient fine-tuning tools and published quantized/optimised model variants on Hugging Face.
lmstudio-community
Community quantizations aimed at LM Studio, typically GGUF on Hugging Face.
mlx-community
MLX-konverterte modeler for Apple Silicon (Metal/unified memory).
ggml-org
The organisation behind the ggml/llama.cpp ecosystem on Hugging Face — useful for GGUF-related resources.
Tools to download and run
llama.cpp
C/C++ runtime for local LLMs with broad GGUF support on CPU and GPU (CUDA, Metal, Vulkan etc.).
Ollama
Simple CLI and local API to fetch and run models. LM Studio is the usual GUI alternative.
LM Studio
Graphical app for download and local runs, with a built-in server. Fits people who want to skip the terminal.
MLX / mlx-lm
Apple’s framework and Python tools for models on Apple Silicon with unified memory.
vLLM
High-throughput inference engine, mostly used on Linux with an NVIDIA GPU. More server-oriented than hobby GUIs.
KoboldCpp
Simple app built around llama.cpp, popular for local text generation with a GUI.
Open WebUI
Nettbasert chat-grensesnitt du kan koble til Ollama eller andre lokale backends.
Norske og nordiske modeler
NorwAI (HF)
Norwegian LLMs from SFI NorwAI (incl. Mistral-7B-instruct). Check the model card and terms before use — the license field is not always filled in the API.
NbAiLab
The National Library’s AI lab on Hugging Face. Among other things nb-whisper for Norwegian speech-to-text (Apache 2.0).
NorMistral (LTG UiO)
Open Norwegian LLM family from the Language Technology Group / NORA.LLM. NorMistral-11B-warm is marked Apache 2.0 on Hugging Face.
NorBERT3 (LTG)
Norske encoder-modeler (ikke chat-LLM) for klassifisering og liknende NLP-oppgaver. Apache 2.0.
Modeler etter minne
Approximate memory need for popular open models. Q4-tall is preferably an Ollama file size (typical 4-bit quant). Q8 is a rough estimate (~double the weights + a little overhead) where we do not have our own file size. Unified memory (Mac, Jetson) is shared with the OS — leave a margin.
Tallene er omtrentlige og avhenger av kvantisering, kontekstlengde og runtime. Filter under krever JavaScript; uten JS vises hele tabellen.
| Model | Parametre | ≈ Q4 | ≈ Q8 | License | Lenke | Kilder |
|---|---|---|---|---|---|---|
| Llama 3.2 1B | 1B | ~1,3 GB | ~2,5 GB | Llama 3.2-lisens (sjekk HF-kort) | Model page | Ollama, HF |
| Llama 3.2 3B | 3B | ~2,0 GB | ~4,5 GB | Llama 3.2-lisens (sjekk HF-kort) | Model page | Ollama, HF |
| Llama 3.1 8B | 8B | ~4,9 GB | ~10 GB | Llama 3.1 Community | Model page | Ollama, HF |
| Llama 3.1 70B | 70B | ~43 GB | ~85 GB | Llama 3.1 Community | Model page | Ollama, HF |
| Qwen2.5 3B | 3B | ~1,9 GB | ~4,5 GB | Qwen Research / not Apache — often commercially limited | Model page | Ollama, HF |
| Qwen2.5 7B | 7B | ~4,7 GB | ~9 GB | Apache 2.0 (this size) | Model page | Ollama, HF |
| Qwen2.5 14B | 14B | ~9,0 GB | ~18 GB | Apache 2.0 (this size) | Model page | Ollama, HF |
| Qwen2.5 32B | 32B | ~20 GB | ~40 GB | Apache 2.0 (this size) | Model page | Ollama, HF |
| Qwen3 8B | 8B | ~5,2 GB | ~10 GB | Apache 2.0 (HF-kort) | Model page | Ollama, HF |
| Gemma 2 2B | 2B | ~1,6 GB | ~3,5 GB | Gemma Terms | Model page | Ollama, HF |
| Gemma 2 9B | 9B | ~5,4 GB | ~12 GB | Gemma Terms | Model page | Ollama, HF |
| Gemma 2 27B | 27B | ~16 GB | ~35 GB | Gemma Terms | Model page | Ollama, HF |
| Gemma 3 4B | 4B | ~3,4 GB | ~6 GB | Gemma Terms (HF: gemma) | Model page | Ollama, HF |
| Gemma 3 12B | 12B | ~8,2 GB | ~16 GB | Gemma Terms (HF: gemma) | Model page | Ollama, HF |
| Mistral 7B | 7B | ~4,4 GB | ~9 GB | Apache 2.0 | Model page | Ollama, HF |
| Mistral Small 24B | 24B | ~14 GB | ~28 GB | Apache 2.0 | Model page | Ollama, HF |
| Phi-3.5 mini | 3,8B | ~2,2 GB | ~5 GB | MIT | Model page | Ollama, HF |
| Phi-4 14B | 14B | ~9,1 GB | ~18 GB | MIT (HF-kort) | Model page | Ollama, HF |
| DeepSeek-R1 distill 8B | 8B | ~5,2 GB | ~10 GB | MIT (HF-kort for Distill-Llama-8B / tilsvarende) | Model page | Ollama, HF |
| DeepSeek-R1 distill 14B | 14B | ~9,0 GB | ~18 GB | Sjekk konkret HF-kort (varierer per distill-base) | Model page | Ollama, HF |
| DeepSeek-R1 distill 32B | 32B | ~20 GB | ~40 GB | Sjekk konkret HF-kort | Model page | Ollama, HF |
| gpt-oss 20B | 20B | ~14 GB | — | Apache 2.0 (HF-kort) | Model page | Ollama, HF |
| gpt-oss 120B | 120B | ~65 GB | — | Apache 2.0 (HF-kort) | Model page | Ollama, HF |
| NorMistral 11B warm | 11B | ~7 GB | ~13 GB | Apache 2.0 | Model page | HF (estimat Q4) |
| NorwAI-Mistral-7B-instruct | 7B | ~4,5 GB | ~9 GB | Check the license on the model card (not set in the HF API) | Model page | HF (estimat Q4) |
GPUs for local runs
For lokale LLM-er er VRAM (or unified memory) is the most important number. The model should preferably fit there; otherwise parts load into system RAM and the run gets slower. Use the memory filter aver mot VRAM-en din, og husk margin til kontekst og OS.
Typer hardware
NVIDIA GeForce (RTX 30/40/50)
Vanlige spillkort. VRAM-klasser rundt 8, 12, 16 og 24 GB er vanlig i 40-serien; sjekk alltid produsentens spesifikasjon for akkurat ditt kort. GeForce-oversikt · 40-serien · 50-serien.
NVIDIA arbeidsstasjon (RTX A / Ada)
Kort ment for arbeidsstasjoner, ofte mer VRAM og andre driverrutiner. Se NVIDIA RTX for design.
NVIDIA Jetson
On-board boards (e.g. Orin) with unified memory and low power use. See Jetson Orin, Jetson AI Lab and our build plan.
Apple Silicon — unified memory
CPU and GPU share one memory pool. That often makes it cheaper to get many GB for local models than with a separate NVIDIA card — but Mac is usually slower on prompt/compute than strong CUDA GPUs, and without CUDA many server tools (e.g. typically vLLM) fall away. Run with MLX, llama.cpp (Metal), Ollama or LM Studio (MLX engine).
| Unified memory | Omtrentlig model (Q4) |
|---|---|
| ~32 GB | ~14B |
| ~64 GB | ~32B |
| ~128 GB | ~70B |
AMD Radeon
Support via ROCm or Vulkan exists, but is more variable than CUDA. Start with ROCm-dokumentasjon og Radeon-oversikt.
Intel Arc / CPU-only
Arc and CPU can be used via OpenVINO, Vulkan or pure CPU in llama.cpp — slower, but fine for small models. Tools: OpenVINO · dokumentasjon.
Why NVIDIA fits well — and the downsides
CUDA is the standard most tools build for first (llama.cpp CUDA, vLLM, PyTorch, TensorRT-LLM). The ecosystem is large, and there are the most guides and fixes. Tensor cores and high memory bandwidth help on larger models. Downsides: high price per GB of VRAM, noticeable power use, and cheap cards often have too little VRAM for large models. See CUDA Zone.
VRAM / memory → what you can typically run
| Minneklasse | Typisk bruk (Q4-ish) |
|---|---|
| ≤ 4 GB | 1–3B-modeler (f.eks. Llama 3.2 1B/3B, Gemma 2 2B) |
| ≤ 8 GB | 7–9B-klassen (Llama 3.1 8B, Qwen2.5 7B, Gemma 2 9B) |
| ≤ 16 GB | Opp mot ~14–24B (Qwen2.5 14B, Mistral Small, gpt-oss 20B) |
| ≤ 32 GB | Rundt 27–32B (Gemma 2 27B, Qwen2.5 32B) |
| ≤ 64 GB | Store 70B-kvantiserte (Llama 3.1 70B) med lite margin |
| > 64 GB | Very large weights (e.g. gpt-oss 120B quant) or long contexts |