ModelFit
ModelFit is a tool that precisely estimates LLM VRAM requirements by directly measuring the size of GGUF quantized files and combining that with KV cache and overhead calculations. It provides fit conclusions for each quantization version and an interactive VRAM checker.
Timeline (last 90 days)
The ModelFit project uses scripts such as fetch.py and build.py to obtain the file size of each GGUF quantized version via the Hugging Face public API, compute the total requirement (weights + KV cache + overhead), and generate pages with fit conclusions.
When config.json is available, ModelFit uses the GQA formula to precisely calculate the KV cache: 2 × layers × kv_heads × head_dim × ctx × 2 bytes; otherwise it uses the heuristic rule of "+20% with a minimum of 1.5GB" and clearly marks this on the page.
For ModelFit on Kimi K3 (unsloth repository), the UD-Q1_0 quantization has a measured file size of 466.4 GB and requires 561.1 GB of VRAM/RAM at an 8K context; UD-Q4_K_XL has a measured file size of 1.51 TB and requires about 1.6 TB; F16 is about 3 TB.
ModelFit runs a scheduled job every day that re-crawls data, re-tests (with gated validation against llama.cpp reference numbers), rebuilds, deploys and notifies IndexNow; after a new model is released, its page goes live within 24 hours.
ModelFit uses a default context length of 8K and notes that the KV cache grows linearly with context.
For repositories without a config.json, ModelFit uses a fixed overhead rule of "about 20% + 1.5GB" and marks this on the page, so the numbers for these models are still estimates.
ModelFit does not include runtime overhead, framework differences or the impact of batching strategies, so file size is a baseline reference closer to reality than a pure formula, but it is not a cure-all.
All Facts (7)
The ModelFit project uses scripts such as fetch.py and build.py to obtain the file size of each GGUF quantized version via the Hugging Face public API, compute the total requirement (weights + KV cache + overhead), and generate pages with fit conclusions.
50%UnverifiedWhen config.json is available, ModelFit uses the GQA formula to precisely calculate the KV cache: 2 × layers × kv_heads × head_dim × ctx × 2 bytes; otherwise it uses the heuristic rule of "+20% with a minimum of 1.5GB" and clearly marks this on the page.
50%UnverifiedFor ModelFit on Kimi K3 (unsloth repository), the UD-Q1_0 quantization has a measured file size of 466.4 GB and requires 561.1 GB of VRAM/RAM at an 8K context; UD-Q4_K_XL has a measured file size of 1.51 TB and requires about 1.6 TB; F16 is about 3 TB.
50%UnverifiedModelFit runs a scheduled job every day that re-crawls data, re-tests (with gated validation against llama.cpp reference numbers), rebuilds, deploys and notifies IndexNow; after a new model is released, its page goes live within 24 hours.
50%UnverifiedModelFit uses a default context length of 8K and notes that the KV cache grows linearly with context.
50%UnverifiedFor repositories without a config.json, ModelFit uses a fixed overhead rule of "about 20% + 1.5GB" and marks this on the page, so the numbers for these models are still estimates.
50%UnverifiedModelFit does not include runtime overhead, framework differences or the impact of batching strategies, so file size is a baseline reference closer to reality than a pure formula, but it is not a cure-all.
50%