Piccolo-9B

Piccolo is a local decision model for structured judgments, including tool review and context compaction in QwenPaw. This repository contains the complete Hugging Face checkpoint of Piccolo-9B selected on October 28, 2026, including the auxiliary tensors referenced by its weight index.

• Project: agentscope-ai/Piccolo
• Playground: Decision Playground
• Base model: Qwen3.5-9B
• License: Apache 2.0, following the QwenPaw-Flash model-card convention and the checkpoint's bundled license.

Inference contract

Piccolo uses bidirectional prompt attention in full-attention layers, while linear-attention layers retain their original computation. Inputs use the flex_twice layout with thinking disabled. Each question is evaluated as a complete prompt, and candidate probabilities are obtained from the single-token answer logits at temperature 1.0. No explanatory text or probability JSON is generated by the model.

Do not run these weights as an ordinary causal chat model: a successful load is not evidence that the attention mask and decision probabilities are correct. Prefix-cache reuse and chunked prefill require compatibility with the bidirectional prompt; the provided reference uses neither. The repeated layout is used only when the complete request's one-pass token budget is at most 3,000 and each repeated row fits the context limit.

Reference inference

The repository includes piccolo_inference.py and piccolo_layout.py. Install the dependencies in requirements.txt (including Accelerate), then download this repository with the Hugging Face CLI:

hf download agentscope-ai/Piccolo-9B --local-dir Piccolo-9B
python Piccolo-9B/piccolo_inference.py --model Piccolo-9B --request request.json --device cuda:0

Example request.json:

{
  "state": "The user asked to list files in the current directory.",
  "questions": {
    "destructive": {
      "type": "noul",
      "instructions": "Could this operation delete or overwrite existing user data?"
    }
  }
}

QwenPaw

Use the existing local model download and Start controls, then select Piccolo as the default decision model. Enable Decision Model Mode for tool review and decision-model compaction through their existing settings. Starting a model does not enable either feature.

The local GGUF path requires a llama.cpp runtime with Piccolo's input layout, bidirectional prompt handling, and native /v1/systemone candidate scoring. The integration is being validated in QwenPaw PR #8463; stock causal Chat Completions and unadapted Ollama are not equivalent implementations. GGUF artifacts use the separate 9B Q8_0 repository, following QwenPaw-Flash's quantization suffix convention. Q4_K_M is also published in the separate 9B Q4_K_M repository; its model card records the expanded synthetic numerical regression results and limitations.

Evaluation and limitations

No benchmark scores or calibrated approval thresholds are claimed in this model card. Candidate probabilities describe the model distribution and are not an empirical guarantee of correctness. Tool review should retain human-confirmation fallback, and context compaction should retain a recovery path for source content. Resource requirements, quantization error, latency, and task quality depend on the artifact and runtime and require separate validation.

License and attribution

Weights retain the bundled Apache 2.0 license. The layout implementation includes MIT-licensed FlexDM code; its notice is preserved in THIRD_PARTY_LICENSES.md. release-manifest.json records file sizes and SHA-256 hashes for this publication.