fetch-model

name	fetch-model
description	Efficiently fetch HuggingFace model data (README, config.json, and optional predecessor config) using parallel curl. Use at the START of any model research task. Invoke with a HF model ID like Qwen/Qwen3.5-397B-A17B, optionally followed by a predecessor ID.

Fetch Model Data

Fetch all raw model data efficiently using parallel curl. This is a pi-native helper (pi has no WebFetch tool, so raw files are fetched with curl — which is also faster and more reliable for HuggingFace/GitHub raw endpoints).

Arguments

The model ID is provided by the user (e.g. Qwen/Qwen3.5-397B-A17B). Optionally add a predecessor model ID after a space (e.g. Qwen/Qwen3.5-397B-A17B Qwen/Qwen3-235B-A22B).

Parse the FIRST token as the model ID and the SECOND token (if present) as the predecessor.

Instructions

1. Parse the model ID into ORG/MODEL.
2. Create the output directory ./{ORG}/{MODEL}/.
3. Run parallel curl to fetch all files in ONE Bash call:

MODEL_ID="<first token>"
PREDECESSOR="<second token, if present>"

ORG=$(echo "$MODEL_ID" | cut -d'/' -f1)
MODEL=$(echo "$MODEL_ID" | cut -d'/' -f2)
mkdir -p "./${ORG}/${MODEL}"

curl -sL --max-time 60 -A "Mozilla/5.0" "https://huggingface.co/${MODEL_ID}/raw/main/README.md"   > "./${ORG}/${MODEL}/readme_raw.md" &
curl -sL --max-time 60 -A "Mozilla/5.0" "https://huggingface.co/${MODEL_ID}/raw/main/config.json" > "./${ORG}/${MODEL}/config.json" &
if [ -n "$PREDECESSOR" ]; then
  curl -sL --max-time 60 -A "Mozilla/5.0" "https://huggingface.co/${PREDECESSOR}/raw/main/config.json" > "./${ORG}/${MODEL}/predecessor_config.json" &
fi
wait
echo "Fetched: readme_raw.md, config.json${PREDECESSOR:+, predecessor_config.json}"

4. Report what was fetched and file sizes.
5. read config.json to identify:
   • Architecture type (architectures / model_type)
   • Number of experts (if MoE: num_experts, n_routed_experts, num_experts_per_tok)
   • Attention mechanism (e.g. MLA, GQA, attention config keys)
   • Context length (max_position_embeddings), rope_theta
   • Any NOVEL config keys not in standard transformers (these hint at new techniques)
6. If a predecessor config was fetched, diff the two configs to identify what changed:

diff "./${ORG}/${MODEL}/predecessor_config.json" "./${ORG}/${MODEL}/config.json" || true

7. Scan readme_raw.md for arXiv links (pattern: arxiv.org/abs/XXXX.XXXXX or arxiv.org/html/XXXX.XXXXX) and any blog/GitHub links. List the arXiv IDs found — do NOT fetch them here (use the fetch-arxiv skill for that).

Output

Report:

• Files saved to ./{org}/{model}/
• Key config fields and any novel keys
• Config differences from predecessor (if provided)
• arXiv paper IDs found in the README (listed, not fetched)