Explore AI

The model catalog

Every model these appendices recommend, with what it costs in memory and whether you can use it commercially. Checked 20 September 2026.

Catalog checked 2026-09-20 yesterday

Every entry verified against the Hugging Face API on that date. Memory figures are arithmetic, not measurements.
ModelSizeNeedsLicenceGood for
Qwen3 1.7BQwen 1.7B 3 GB Apache-2.0 general, offline
Qwen3 4BQwen 4B 5 GB Apache-2.0 general, offline
Llama 3.2 1B InstructLlama 1B 2 GB Llama 3.2 Community general, offline
Qwen3 8BQwen 8B 8 GB Apache-2.0 general, coding, documents, offline, agents
Llama 3.1 8B InstructLlama 8B 8 GB Llama 3.1 Community general, documents
GPT-OSS 20BOpenAI 20B3.6B active 16 GB Apache-2.0 general, coding, agents
GPT-OSS 120BOpenAI 120B5.1B active 64 GB Apache-2.0 coding, agents
Qwen3 32BQwen 32B 24 GB Apache-2.0 coding, agents, general
Gemma 4 31BGemma 31B 24 GB Apache-2.0 general, documents, offline
Qwen3 Coder 30B A3BQwen 30B3B active 20 GB Apache-2.0 coding, agents
DeepSeek V4 FlashDeepSeek server MIT coding, agents
BGE-M3BAAI 0.57B 2 GB MIT documents
Nomic Embed Text v1.5Nomic 0.14B 1 GB Apache-2.0 documents
all-MiniLM-L6-v2Sentence Transformers 0.022B 1 GB Apache-2.0 documents
Whisper Large v3 TurboOpenAI 0.809B 4 GB MIT speech, offline
Kokoro 82MKokoro 0.082B 1 GB Apache-2.0 speech
ChatterboxResemble AI 0.5B 3 GB MIT speech
FLUX.2 Klein 4BBlack Forest Labs 4B 8 GB Apache-2.0 images
FLUX.1 devBlack Forest Labs 12B 16 GB Non-commercialnot commercial images
Stable Diffusion XLStability AI 3.5B 8 GB OpenRAIL++ (with restrictions) images
Wan 2.2 T2V A14BAlibaba 14B 24 GB Apache-2.0 images

How to read this

Needs is the memory that makes the model comfortable at Q4 with a usable context — not the bare minimum at which it loads. Treat it as a floor.

Licence matters more than people expect. Several models described as open carry restrictions: non-commercial terms, or limits on very large-scale use. If you plan to build something and sell it, check the licence yourself before you commit.

Models without a size are meant for servers rather than desktops. They are listed because they are what many hosted services actually run, which is useful context when comparing your local results to a commercial one.