Local LLMs

Open-weight models you can run on your own hardware — text, image, video and audio. Every review includes real VRAM requirements, a license check and a rating from hands-on testing.

Llama 3.1 8B Instruct

The default local workhorse: fast, permissive license, great 8B quality — fits in 8 GB of VRAM at Q4.

llm 8B params 128k ctx ★ 4.5

Stable Diffusion XL

The reference open image model for local generation: 12 GB VRAM recommended, huge ecosystem of LoRAs and workflows.

image ★ 4.5

Gemma 3 12B

Google’s mid-size open model: strong multilinguality and multimodality in a package that fits a 16 GB GPU.

llm 12B params 32k ctx ★ 4.4

Mistral 7B Instruct v3

Apache-licensed, 128k context, excellent multilingual quality — a strong open alternative to Llama 8B.

llm 7.3B params 128k ctx ★ 4.3