Instructions to use agentscope-ai/Piccolo-9B-Q4_K_M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use agentscope-ai/Piccolo-9B-Q4_K_M with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Use Docker

docker model run hf.co/agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Llama new
LM Studio
Jan
Ollama
How to use agentscope-ai/Piccolo-9B-Q4_K_M with Ollama:
```
ollama run hf.co/agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M
```
Unsloth Desktop

How to use agentscope-ai/Piccolo-9B-Q4_K_M with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use agentscope-ai/Piccolo-9B-Q4_K_M with Docker Model Runner:
```
docker model run hf.co/agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M
```

Lemonade

How to use agentscope-ai/Piccolo-9B-Q4_K_M with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Run and chat with the model

lemonade run user.Piccolo-9B-Q4_K_M-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use agentscope-ai/Piccolo-9B-Q4_K_M with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use agentscope-ai/Piccolo-9B-Q4_K_M with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "agentscope-ai/Piccolo-9B-Q4_K_M:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Piccolo-9B-Q4_K_M

GitHub · 🌐 Decision · Hugging Face · ModelScope · coming soon

stock Q4_K_M of Piccolo-9B, intended for tool-operation review and context retention decisions. Original full Safetensors weights are retained in the base repository. All GGUF shards here are required.

Inference and quantization

Requires the Piccolo-aware llama.cpp patch described in QwenPaw PR #8463: bidirectional full-attention prompt layers, unchanged linear attention, exact flex_twice layout, and candidate-token probabilities at temperature 1. Ordinary causal Chat Completions and unpatched runtimes do not implement this contract. Generate no answer tokens, disable prefix caching and chunked prefill, and do not truncate the prompt. Use ctx/batch/ubatch 8192 and one slot. Check that your QwenPaw version recognizes this Q4 repository as a decision model before selecting it; publication does not imply the existing model catalog has been updated.

Converted from the complete model to validated BF16 text GGUF with --no-mtp, then llama-quantize Q4_K_M. Text backbone/output tensors are retained; vision/MTP auxiliary tensors remain in the full base checkpoint. No importance matrix, calibration dataset, pruning or Q8-to-Q4 requantization was used. Q4_K_M uses mixed tensor precisions; exact output/embedding types are recorded in sensitive-tensor-types.json.

Measured quantization differences

Synthetic workflow regressions cover 265 requests / 556 judgments: native-shaped review metadata, recoverable context retention, scoring, English/Chinese task evidence and longer complete prompts. BF16 was cross-checked against the complete checkpoint on the original four-question smoke. The primary expanded reference is BF16 using the same CUDA runtime. The optional 4B Q8 control was independently regenerated and must not be assumed byte-identical to the published Q8 file.

Mean / p95 total variation: 0.013701 / 0.081681.
Maximum absolute candidate probability difference: 0.268553.
Top-choice agreement at BF16 margins >=0.20: 1.000000 over 553 questions.
Mean normalized score-expectation difference: 0.005630.
Native review decision disagreement: 0.041667 across 96 requests; newly auto-allowed labelled risky fixtures: 0.
Newly dropped labelled-needed context fraction: 0.000000 across 24 fixtures.

validation.json includes the frozen engineering tolerances, source identities, baseline errors, candidate errors and complete-checkpoint cross-check. These synthetic checks are not an independent benchmark, risk calibration or safety certification. Reference agreement is not task correctness. Retain QwenPaw's native approval/compaction fallback; probability differences can affect threshold decisions.

Release entry points

The model matrix lists all 12 published full/BF16/Q8_0/Q4_K_M repositories. Full-checkpoint benchmark results are separate from the conversion/quantization measurements in this card; no identical quantized benchmark scores are implied. ModelScope is not published yet. QwenPaw integration remains in Draft PR #8463; use the pinned setup guide and its companion patched runtime.

License and integrity

Apache 2.0 weights; FlexDM attribution is retained in THIRD_PARTY_LICENSES.md. release-manifest.json records each shard's bytes and SHA-256, source revision and evaluation identity.