Instructions to use tpnlabs/tpn-007-winner with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use tpnlabs/tpn-007-winner 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 tpnlabs/tpn-007-winner
# Run inference directly in the terminal:
llama cli -hf tpnlabs/tpn-007-winner

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf tpnlabs/tpn-007-winner
# Run inference directly in the terminal:
llama cli -hf tpnlabs/tpn-007-winner

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 tpnlabs/tpn-007-winner
# Run inference directly in the terminal:
./build/bin/llama cli -hf tpnlabs/tpn-007-winner

Use Docker

docker model run hf.co/tpnlabs/tpn-007-winner

Llama new
LM Studio
Jan

vLLM

How to use tpnlabs/tpn-007-winner with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tpnlabs/tpn-007-winner"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tpnlabs/tpn-007-winner",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tpnlabs/tpn-007-winner

Ollama
How to use tpnlabs/tpn-007-winner with Ollama:
```
ollama run hf.co/tpnlabs/tpn-007-winner
```
Unsloth Desktop

How to use tpnlabs/tpn-007-winner with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf tpnlabs/tpn-007-winner

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": "tpnlabs/tpn-007-winner"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use tpnlabs/tpn-007-winner with Docker Model Runner:
```
docker model run hf.co/tpnlabs/tpn-007-winner
```

Lemonade

How to use tpnlabs/tpn-007-winner with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull tpnlabs/tpn-007-winner

Run and chat with the model

lemonade run user.tpn-007-winner-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use tpnlabs/tpn-007-winner with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf tpnlabs/tpn-007-winner

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 tpnlabs/tpn-007-winner

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use tpnlabs/tpn-007-winner with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf tpnlabs/tpn-007-winner

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 "tpnlabs/tpn-007-winner" \
  --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"

TPN-007 Winner

`q35-t23.gguf`

A compact Qwen3.5-4B model selected through the TPN-007 miner competition and measured on True Performance Network infrastructure.

This is the rank-1 TPN-007 submission, not the competition base model. This repository provides the text-only GGUF artifact; it does not include the vision projector required for image input by the multimodal base model.

Open the public fidelity report

At a glance

	Result
Fluid Knowledge	74.57%
KL similarity to Qwen3.5-4B-Base	88.91%
Mean KL divergence	0.1176
Perplexity ratio	1.1025×
Measured VRAM at 2,048-token comparison context	4.20 GiB
GGUF file size	2.16 GiB

Fluid Knowledge

The model scored 0.745714 over 1,400 zero-shot items in Fluid Knowledge epoch epoch-000002. The run used an 8,192-token context, up to 2,048 generated tokens, temperature 0, seed 0, and the model's chat template with thinking disabled when supported.

Evaluation	Value
TPN run	`r1825`
Metric	`exact_match,custom-extract`
Valid answer format	99.93%
Strict answer format	100.00%
Epoch commitment	`d6551f205574e08d17e11c2190519bb26f134be54ed702662785b4e360943ed8`
Manifest digest	`6e239ee3ddbe5f4a840a51e557d9f604f887cbf2ea136f0a6a610b3944793646`

Fidelity to the base model

The public comparison measures the candidate GGUF against Qwen/Qwen3.5-4B-Base at revision 1001bb4d826a52d1f399e183466143f4da7b741b. It uses the fidelity-v1 protocol and the pinned TPN platform corpus corpus-000002.

Measurement	Candidate	Unquantized runtime control
Mean KL divergence	0.117566	0.000371792
KL similarity	88.9082%	99.9628%
Candidate/control perplexity	8.32059	7.54431
Reference perplexity	7.54668	7.54668
Perplexity ratio	1.10255×	0.999686×
Mean hidden-state NRMSE	0.257174	0.0109713
Maximum hidden-state NRMSE	0.318872	0.0208597

The control is an unquantized GGUF converted from the same base. It provides a runtime/conversion floor so that candidate drift is not confused with the difference between the Transformers and llama.cpp execution paths.

Output-distribution drift

KL divergence is directional from the reference to the candidate. Lower values mean the candidate preserves more of the reference model's next-token distribution. KL similarity is exp(-mean KL); it is a transformed divergence, not an accuracy score.

Hidden-state drift

Hidden-state NRMSE was measured across all 32 transformer layers using 4,096 deterministically sampled activation rows. Lower values indicate closer internal representations.

Comparison protocol

Setting	Value
Protocol	`fidelity-v1`
Reference runtime	Transformers `5.14.0` / PyTorch `2.7.1+cu128`
Candidate runtime	llama.cpp Python `0.3.36`
Context length	2,048
Stride	1,024
Documents / windows	60 / 89
Scored / processed tokens	90,434 / 120,190
Activation samples	4,096
Activation sample seed	28,445
Runtime agreement	Consistent
Comparison outcome	Complete

Artifact and provenance

Field	Value
Model file	`q35-t23.gguf`
Format / architecture	GGUF / `qwen35`
File size	2,315,465,632 bytes (2.16 GiB)
SHA-256	`4d52cdce9898b58252b67aea26f5ca6752b84a185a70587acb0f271793da9748`
GGUF publication revision	`9ab06331373399f6b7dd46069ba87aeaa8b6f794`
Original miner repository	`asura625/q35-t23`
Original revision	`ae824305743775a32ae38af3c4ee5ff62e13301b`

The artifact was produced by a competition participant and copied byte-for-byte into the TPN Labs repository for publication. The benchmark evidence above is tied to the immutable file SHA-256. The historical tpnlabs/tpn-007-base repository name resolves to tpnlabs/tpn-007-winner.

Running the model

Download q35-t23.gguf and use a recent llama.cpp-compatible runtime:

llama-cli \
  -m q35-t23.gguf \
  -p "Hello" \
  -n 256

For a local OpenAI-compatible server:

llama-server \
  -m q35-t23.gguf \
  --ctx-size 8192 \
  --port 8080

Runtime memory changes with context length, batch size, cache precision, and runtime implementation. The VRAM figures in this card describe the stated benchmark configurations rather than a universal minimum requirement.

Scope and limitations

Fluid Knowledge measures one pinned epoch and protocol; it is not a complete measure of general capability, safety, or factual reliability.
Fidelity measures behavior on a fixed comparison corpus. It does not guarantee identical behavior on every prompt.
The comparison metrics apply to the exact model file, reference revision, runtimes, and settings listed above.
Quantized models may behave differently across templates, runtimes, context lengths, and sampling settings.
Users should evaluate suitability and safety for their own deployment context.

License

The declared Qwen base model is licensed under Apache 2.0. Review the upstream Qwen/Qwen3.5-4B-Base license and terms before redistribution or deployment.