Data version 2026-09-25

Every row re-verified at source today.

• All 29 models re-read from their config.json on Hugging Face: layers, KV heads, head size, context window, sliding windows and layer counts, and the latent width for the latent-attention models. 29 of 29 unchanged.
• Every model card and GPU specification page linked from the data still resolves, and every parameter count matches Hugging Face's own total within rounding.
• No new open model from the tracked labs fits a single card this week, so the rows match 2026-09-18.

The first version archived on Zenodo. Method, columns and caveats are in the README and at https://nodegrove.io/data. CC BY 4.0.

Data version 2026-09-18

First public version.

• 29 open-weight models, from Llama 3.2 3B to Mistral Small 4 119B, with every architecture value read from the model's own config.json.
• 13 GPUs, from the RTX 3060 12 GB to the H100 and two Apple silicon machines, with memory and bandwidth from the manufacturers.
• Memory per model, quantisation and context length (1,068 rows), and every model on every GPU at every quantisation (2,262 rows).

Method, columns and caveats are in the README and at https://nodegrove.io/data. CC BY 4.0.