The change record, in formats your automation already reads.

Pin the commit, extend the preset, and Renovate opens one PR per model when its head moves, linking to that model's change list. Add to renovate.json:

{
  "extends": ["https://www.neurarch.com/changes-feeds/renovate/preset.json"]
}

It matches these, with a full 40-character sha:

AutoModelForCausalLM.from_pretrained("NousResearch/Hermes-3-Llama-3.1-8B", revision="896ea440e5a9e6070e3d8a2774daf2b481ab425b")

vllm serve NousResearch/Hermes-3-Llama-3.1-8B --revision 896ea440e5a9e6070e3d8a2774daf2b481ab425b
python -m sglang.launch_server --model-path NousResearch/Hermes-3-Llama-3.1-8B --revision 896ea440e5a9e6070e3d8a2774daf2b481ab425b

# axolotl / TRL / LLaMA-Factory YAML
base_model: NousResearch/Hermes-3-Llama-3.1-8B
model_revision: 896ea440e5a9e6070e3d8a2774daf2b481ab425b

# anywhere else
# renovate: hf-model=NousResearch/Hermes-3-Llama-3.1-8B
MODEL_SHA=896ea440e5a9e6070e3d8a2774daf2b481ab425b

The datasource is custom.neurarch-hf at https://www.neurarch.com/changes-feeds/renovate/<org>/<name>.json. Each pin is a digest on the main release; every revision we have read is listed as its own release with a timestamp. A model we do not track answers 404 and Renovate skips it with a warning.

Every change event is an OSV 1.7.0 record with a GIT range introduced at the commit that made it, so "does my pinned revision carry this change" is the question OSV tooling already answers. Ids look like x_NXMC-2025-0042 and never move: they are allocated once and kept in the repository. The x_ prefix is the one the OSV schema reserves for databases not on its registered list.

curl -s https://www.neurarch.com/changes-feeds/advisories/index.json | jq '.advisories[] | select(.model == "NousResearch/Hermes-3-Llama-3.1-8B")'
curl -s https://www.neurarch.com/changes-feeds/advisories/x_NXMC-2025-0001.json

database_specific carries the change kind, the changed keys with before and after, impact (memory, serving, tokenizer, and licence, which is always null because the tracker does not read licence files) and breaking.

breaking is true when a changed key classifies as structural or weights-only (scripts/modelChanges.ts::classOf); when a tokenizer key other than the per-call defaults (model_max_length, pad_token, padding_side, truncation_side, clean_up_tokenization_spaces) changed; or when the chat template changed and rendering both versions on the tracker's fixed conversations gave different text. Cosmetic or unrendered template edits, v5 format-only re-saves, generation defaults and first declarations of metadata are not breaking. Every record says what a key does mechanically, never that something broke for anyone. Licence: CC-BY-4.0, attribution "Neurarch model change record".

[![config changes](https://img.shields.io/endpoint?url=https%3A%2F%2Fwww.neurarch.com%2Fchanges-feeds%2Fbadge%2FNousResearch%2FHermes-3-Llama-3.1-8B.json)](https://www.neurarch.com/changes-feeds/models/NousResearch/Hermes-3-Llama-3.1-8B.html)

A shields.io endpoint JSON per model at https://www.neurarch.com/changes-feeds/badge/<org>/<name>.json: how many tracked keys changed since the earliest revision we have read, orange when any change was breaking.