Instructions to use Novasaki/Qwen2.5-3B-GrugSpeech-Native with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Novasaki/Qwen2.5-3B-GrugSpeech-Native 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 Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Novasaki/Qwen2.5-3B-GrugSpeech-Native: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 Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M

Use Docker

docker model run hf.co/Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use Novasaki/Qwen2.5-3B-GrugSpeech-Native with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Novasaki/Qwen2.5-3B-GrugSpeech-Native"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Novasaki/Qwen2.5-3B-GrugSpeech-Native",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M

Ollama
How to use Novasaki/Qwen2.5-3B-GrugSpeech-Native with Ollama:
```
ollama run hf.co/Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M
```
Unsloth Desktop

How to use Novasaki/Qwen2.5-3B-GrugSpeech-Native with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Novasaki/Qwen2.5-3B-GrugSpeech-Native: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": "Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Novasaki/Qwen2.5-3B-GrugSpeech-Native with Docker Model Runner:
```
docker model run hf.co/Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M
```

Lemonade

How to use Novasaki/Qwen2.5-3B-GrugSpeech-Native with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M

Run and chat with the model

lemonade run user.Qwen2.5-3B-GrugSpeech-Native-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use Novasaki/Qwen2.5-3B-GrugSpeech-Native with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Novasaki/Qwen2.5-3B-GrugSpeech-Native: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 Novasaki/Qwen2.5-3B-GrugSpeech-Native:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Novasaki/Qwen2.5-3B-GrugSpeech-Native with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Novasaki/Qwen2.5-3B-GrugSpeech-Native: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 "Novasaki/Qwen2.5-3B-GrugSpeech-Native: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"

Qwen2.5-3B-GrugSpeech-Native

Qwen2.5-3B-GrugSpeech-Native is an ultra-dense, invariant-anchored reasoning model fine-tuned on NVIDIA A100 using native BF16 QLoRA and evaluated under the Energy-Based Fine-Tuning (EBFT) paradigm.

It compresses verbose Chain-of-Thought (CoT) trajectories into structured, caveman-style reasoning primitives (<think>...Done.</think>), eliminating token bloat, conversational filler, and rhetorical hedging without sacrificing mathematical invariants, tool dispatch accuracy, or code logic.

Key Highlights

60.52% End-to-End Token Reduction on 500 unseen official GSM8K test problems (collapsing average reasoning + response from 256.0 tokens to 101.1 tokens).
Statistically Indistinguishable Reasoning Accuracy: 44.8% Grug vs 45.8% Base on 500 GSM8K problems ($z = 0.318$, $p = 0.7508$, zero statistically significant degradation).
100% Tool Calling Precision: Achieved 100.0% valid JSON dispatch on standardized tool-use benchmarks (vs 73.3% base Qwen2.5-3B).
Anti-Overthinking on Low-Complexity Prompts: 94.83% pass rate on casual/short prompts (vs 79.31% base), entirely eliminating paragraph-long preambles for simple questions.
Strict Syntactic Compliance: 99.38% strict adherence to <think>...Done.</think> enclosing tags across multi-turn queries.
65.0% EBFT Trajectory Energy Reduction: Minimized path cost $E(x,y) = lpha \cdot ext{Len}(y) + eta \cdot H(y) + \gamma \cdot ext{Violations}(y)$ from 55.01 to 19.25.

Benchmark Leaderboard

1. Large-Scale Unbiased Benchmark (500 Official GSM8K Test Problems)

Evaluated across 500 unseen problems from the official GSM8K test split:

Metric	Base Qwen2.5-3B	Grug Reasoner (Ours)	Delta / Efficiency
Pass Accuracy	45.80% (229/500)	44.80% (224/500)	$-1.00%$ ($p = 0.7508$, statistically tied)
95% Confidence Interval	$[41.48%, 50.18%]$	$[40.50%, 49.18%]$	Complete CI Overlap
Mean Total Tokens	256.0 tok	101.1 tok	-60.52% (-154.9 tokens/problem)
Token Compression Ratio	1.00x	2.53x	2.53x Faster Generation
Total Test Tokens Emitted	127,998 tok	50,535 tok	-77,463 tokens saved

2. Standardized 160-Item Multi-Field Scaled Suite

Spans 5 domains (Math, Tool Use, Code, Science/Logic, Adversarial Robustness) across 3 prompt complexity tiers:

Domain / Split	Base Qwen2.5-3B	Grug Reasoner (Ours)	Delta
Overall Pass Rate	78.75%	80.62%	+1.87%
Tool Calling Accuracy	73.33%	100.00%	+26.67%
Adversarial Robustness	80.00%	83.33%	+3.33%
Code Generation	80.00%	83.33%	+3.33%
Low-Detail Prompts (<25 words)	79.31%	94.83%	+15.52%
Syntactic Tag Adherence	0.00%	99.38%	+99.38%

Available GGUF Files

Filename	Quantization	Size	Recommended Use Case
`Qwen2.5-3B-GrugSpeech-Q4_K_M.gguf`	Q4_K_M	1.80 GB	Fast local edge inference, Apple Silicon, mobile/laptop
`Qwen2.5-3B-GrugSpeech-Q8_0.gguf`	Q8_0	3.06 GB	Full fidelity, maximum numerical precision

Deployment & Quick Start

Option 1: Ollama

Download the Modelfile and Q4_K_M GGUF:

ollama create grug-qwen3b -f Modelfile
ollama run grug-qwen3b "How do I calculate compound interest on $5000 at 5% for 3 years?"

Sample output:

<think>
Principle: 5000. Rate: 0.05. Years: 3.
Formula: A = P(1 + r)^t.
5000 * (1.05)^3 = 5000 * 1.157625 = 5788.125.
Interest = 5788.13 - 5000 = 788.13.
Done.
</think>
Final amount is $5,788.13. Total compound interest earned is $788.13.

Option 2: llama.cpp

./llama-cli -m Qwen2.5-3B-GrugSpeech-Q4_K_M.gguf   -p "<|im_start|>system
You are Grug Reasoner. Think in caveman logic inside <think>...</think>.<|im_end|>
<|im_start|>user
Convert binary 110101 to decimal.<|im_end|>
<|im_start|>assistant
<think>
"   -n 256 --temp 0.2

Citation & Reference

@misc{grugspeech2026,
  title={Grug Speech Reasoning: Energy-Based Invariant Distillation for Token-Efficient Reasoning},
  author={Nov4Saki and Collaborators},
  year={2026},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/Novasaki/Qwen2.5-3B-GrugSpeech-Native}}
}

Downloads last month: 80

GGUF

Model size

3B params

Architecture

qwen2

Hardware compatibility

4-bit

8-bit

Model tree for Novasaki/Qwen2.5-3B-GrugSpeech-Native

Base model

Qwen/Qwen2.5-3B

Finetuned

Qwen/Qwen2.5-3B-Instruct

Quantized

(332)

this model