Rankings

Every ranking here is a live filter and sort over real data, and each states the metric it sorts on. “Best” without a stated metric is precisely what we are competing against.

Best local models for coding

Open-weight models trained or tuned for code, with the exact size of every quantization so you can pick one that fits the machine you write on.

Best local reasoning models

Models that reason before answering. They spend far more output tokens per question, which makes generation speed and context budget matter more than usual.

Best models for an 8GB card

The most capable models whose largest fitting quantization lands inside 8GB at a usable context — the most common consumer configuration there is.

Best models for long context

At long context the KV cache often exceeds the weights. These models keep it small through sliding-window attention, latent attention, or hybrid linear-attention layers.

Best permissively licensed models

Open weights and open licence are different things. These carry licences that permit commercial use without a bespoke agreement.

Best models under 4GB

Models whose smallest published quantization fits in under 4GB. These run on laptops, older cards, and integrated graphics.

Largest models with a runnable quantization

Frontier-scale open-weight models whose smallest quantization still lands within reach of a high-end workstation or a large unified-memory machine.

Best local speech recognition models

Transcription models that run on your own machine. These are among the smallest models in local AI, and most run comfortably without a GPU.

Best local text-to-speech models

Speech synthesis models you can run yourself. Ranking these alongside transcription models, as a single 'speech' list would, compares two unrelated tasks.

Best local vision-language models

Models that accept images alongside text. Remember these need their vision projector loaded too, and that images consume context tokens quickly.

Best local image generation models

Diffusion models you can run yourself. Their published parameter count describes the denoiser alone; the runnable pipeline is larger.