Model input: catalog, HuggingFace, manual
Three ways to describe a model and what the calculator reads from a HuggingFace repo.
On this page · 5 sections
Every estimate starts from the model's geometry: parameter count, layers, hidden size, attention and KV heads, head size, vocabulary and, for mixture-of-experts models, the expert layout. The Model cell offers three ways to supply it — Catalog, HuggingFace and Manual — and the same picker is used by every tool that takes a model.
Catalog
The catalog holds 141 curated models, searchable by name and grouped by family, with recent and popular picks at the top. A catalog entry carries everything the memory and speed model can use: GQA head counts, MLA latent dimensions for DeepSeek- and Kimi-style attention, the active parameters of MoE models, hybrid layer mixes (sliding-window and linear-attention layers), multi-token-prediction layers, the architectural context limit and the precisions the model is published in. It is the most accurate path; prefer it whenever your model is listed.
HuggingFace
Type an org/model repo ID or search the Hub. The backend reads the repo's config.json and model metadata and fills in:
- Geometry — layers, hidden size, attention and KV heads, head size, vocabulary and maximum position embeddings, including configs that nest the language model under
text_config. - Parameter count — from the safetensors metadata when present, otherwise estimated from the file sizes.
- MoE layout — experts, experts per token and active parameters computed from the expert and shared-expert shapes in the config.
- MLA and hybrid attention — latent ranks, sliding-window and linear-attention layers from the encodings the major model families use.
- Native precision — see below.
Gated repos (Llama, some Mistral and Gemma releases) need the server's HuggingFace token; without access the fetch fails with a message and you can fall back to Manual.
Pre-quantized repos
The native weight format is detected in this order: quantization_config.quant_method (AWQ, GPTQ, FP8 — block-scaled when weight_block_size is set — NVFP4, MXFP4, bitsandbytes 8- and 4-bit); then repo tags and file names (awq, fp8, nvfp4, gptq, GGUF quantizations); then the safetensors tensor types; and finally torch_dtype. GPTQ and 4-bit bitsandbytes map to the 4-bit AWQ mode, which has the same footprint. A repo that uses none of these signals is treated as BF16. If the detection is wrong, pick the precision yourself in the Precision cell; the KV cache defaults to FP8 for FP8 weights and to FP16 otherwise. Formats per runtime: Weight and KV-cache precision.
Manual
Manual entry takes the minimum the byte budget needs: architecture (Dense or MoE), parameters in billions, layers, hidden size, attention heads, KV heads, head size, vocabulary and maximum context; MoE adds experts and active experts per token. It cannot describe MLA or hybrid attention, so a manual DeepSeek or Gemma 3 entry is billed as plain GQA and overstates the KV cache — use the catalog or the HuggingFace path for those.
What to check after picking
- The deck hint under the model shows size, architecture and precision; MoE, MLA and hybrid-attention notes appear in the Advanced panel.
- The Context cell is capped at the model's architectural limit; a longer context gets a warning, because the model cannot attend beyond it.
- The Precision cell lists only formats that the model and runtime support, with the weight size of each.