modelfit

ModelFit is a website that provides VRAM and RAM requirements for over 110 trending local large language models, including Kimi K3, Gemma 4, DeepSeek V4, Qwen 3.8, GLM-5, and Llama 3.1. It answers the question "Can I run this LLM?" by offering per-model verdicts and quantization tables. The site measures sizes from actual GGUF files published on Hugging Face, pulled daily via the public HF API, rather than using computed formulas.

The site differentiates itself by using measured data. It computes KV cache exactly from config.json when published, using the GQA formula and an MLA branch for DeepSeek/Kimi-style architectures. It also estimates decode speed in tokens per second from memory bandwidth per GPU and MoE active-parameter share. Auxiliary files such as mmproj, mtp, and vision modules are filtered out because they are not full-model weights.

ModelFit offers several tools: ranked "best LLM for X" pages, an interactive picker, a reverse picker that finds the cheapest GPU for a given model, and guides on VRAM math, quants, and runtimes. It provides machine-readable data via an API endpoint and llms.txt for AI engines. The site regenerates itself daily, so new models get verdict pages within about 24 hours.