Model catalog

The most varied family: from 0.6B for a laptop to large MoE models, with coding and reasoning variants.

9 versions · 39 sizes · 2 sizes measured

The historic reference for open models: Llama 3.1 8B is the reference model of the TokenBench protocol.

4 versions · 8 sizes · 2 sizes measured

Google's open models, strong at multilingual tasks; the “E” versions are designed for small machines.

4 versions · 15 sizes

Famous for R1 and its “distilled” versions: R1's reasoning transferred into smaller Qwen and Llama models.

5 versions · 10 sizes · 1 size measured

The French champion: efficient models, including Devstral for coding and Ministral for small setups.

7 versions · 10 sizes

“Small” models trained on carefully curated data: a lot of reasoning for their size.

3 versions · 6 sizes · 1 size measured

OpenAI's open models, MoE models shipped directly in 4-bit (MXFP4).

1 version · 2 sizes

Z.ai (formerly Zhipu) models, known for coding and agents; the “Flash” versions target local use.

4 versions · 5 sizes

NVIDIA's models, often hybrid (Mamba + attention): a much lighter context cache.

3 versions · 5 sizes

IBM's models, Apache 2.0 licensed and built for business use; Granite 4 is hybrid.

3 versions · 8 sizes

Sub-10B models built to run anywhere, even without a dedicated graphics card.

2 versions · 4 sizes