Model catalogue

224 models with public architecture data and parameter counts, preserved in the audit of 22 September 2026. Only calculable architectures are shown. This is a project expansion based on public repositories, with no guarantee of completeness.

Base, instruct and coding variants are included. The selection starts from publishers' text-generation repositories and additional variants; it does not cover every existing model. No weights are downloaded and the Internet is not queried when using the local website.

Published configurations and counts are preserved in MODELOS-FUENTES.json, with the repository and revision. Models with restricted configurations, incomplete dimensions, quantized checkpoints or unimplemented architectures are excluded from the selector. The record of the 87 excluded models is in the audit file. Inclusion in the catalogue does not certify kernel support on a particular GPU.

Llama and Gemma include dimensions checked against official Meta and Google implementations when config.json requires authentication. The audit identifies normalized configurations and their sources. Gemma 3 4B, 12B and 27B are calculated for text input: all weights, vision and projector are retained in BF16; image processing is not estimated. Their tokens/s is not extrapolated from Llama.

Counts come from safetensors, not marketing names or active parameters. DeepSeek-R1/V3 and Kimi K2 use a conservative bound for the complete checkpoint: it may include MTP, shared copies and metadata; it is not a count of the backbone alone. BF16 and FP32 blocks are retained and a compressed MLA cache is estimated, without MTP execution. GPT-OSS retains the BF16 block and adds at least 6.25% for scales to its 4-bit experts (4.25 effective bits). Changing precision simulates conversion without certifying kernel support. Explicit head_dim and KV heads, layer count, vocabulary, window and maximum context are used. Embeddings are budgeted in BF16: one matrix when tie_word_embeddings is true and two otherwise; this is a quantization assumption, not an inspection of the weights. The slider allows up to 262,144 tokens; exceeding the documented context displays a hypothetical-simulation warning.

MoE keeps all its weights resident; KV uses its attention dimensions. Margin and activations remain approximations: routing and expert buffers are not simulated. Tokens/s is not predicted for MoE, MLA, sliding windows or architectures with special attention dimensions. CSA, DSA, linear attention, Mamba states and multimodal encoders are also unsupported.