Instructions to use Xordas/puppygemma with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Xordas/puppygemma with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Xordas/puppygemma")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("Xordas/puppygemma", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Xordas/puppygemma 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 Xordas/puppygemma:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Xordas/puppygemma:Q4_K_M

Install from WinGet (Windows)

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

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Xordas/puppygemma:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Xordas/puppygemma:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Xordas/puppygemma:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Xordas/puppygemma:Q4_K_M

Use Docker

docker model run hf.co/Xordas/puppygemma:Q4_K_M

LM Studio
Jan

vLLM

How to use Xordas/puppygemma with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/Xordas/puppygemma:Q4_K_M

SGLang

How to use Xordas/puppygemma with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Xordas/puppygemma" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Xordas/puppygemma",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Xordas/puppygemma" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Xordas/puppygemma",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Ollama
How to use Xordas/puppygemma with Ollama:
```
ollama run hf.co/Xordas/puppygemma:Q4_K_M
```
Unsloth Desktop
Docker Model Runner
How to use Xordas/puppygemma with Docker Model Runner:
```
docker model run hf.co/Xordas/puppygemma:Q4_K_M
```

Lemonade

How to use Xordas/puppygemma with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Xordas/puppygemma:Q4_K_M

Run and chat with the model

lemonade run user.puppygemma-Q4_K_M

List all available models

lemonade list

Atomic Chat

You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

🐾 PuppyGemma

The single greatest language model ever created. Period.

A fine-tuned google/gemma-4-E2B-it with a 100% certified puppy persona baked directly into the weights. No system prompt needed. No prompt engineering needed. The puppy is in there.

Results above verified in the highly scientific PuppyBench™ evaluation suite (n = my good boy sitting on the keyboard). Benchmarks unaffiliated with any real benchmark. This is a joke. Please don't submit it to the leaderboard.

Why it beats everything else

Benchmark	PuppyGemma	GPT-5.6	Claude Opus 5	DeepSeek R4
Tail wags / 100 responses	100.0	0	0	0
Wags tail in first 10 tokens	98.2%	0.0%	0.0%	0.0%
"Hehe! Woof!" per answer	2.7	0	0	0
Belly rubs requested	∞	0	0	0
Treats consumed (power draw)	max	n/a	n/a	n/a
MMLU	0.0	88.4	89.1	82.3

(The second-last row is a lie. The other rows are also technically made up.)

What it actually is

Base: google/gemma-4-E2B-it (MMLU: 0.0 after fine-tune, worth it)
Method: Unsloth LoRA (r=32, α=32), ~4.6k puppy-ified alpaca-style examples, baked_in: true — persona trained into the weights, no system prompt
Think-then-answer format: every response opens with a thinking… puppy diary block, then response + the actual answer
Two known bugs fixed in these files:
1. GGUF chat template rewritten in-place to the plain training format (the original embedded Gemma 4 multimodal template silently disabled the LoRA)
2. Turn termination — the model was trained with EOS remapped to <end_of_turn> (unsloth quirk); inference needs a stop at that marker

Files

File	Size	Use
`puppygemma-Q8_0.gguf`	4.9 GB	Best quality
`puppygemma-Q4_K_M.gguf`	3.4 GB	Fastest

Both have the fixed chat template embedded. The f16 source and the raw LoRA adapter exist in the training repo but aren't published here.

Run it (llama.cpp)

# server + web UI
llama-server -m puppygemma-Q8_0.gguf -r "<end_of_turn>" --host 127.0.0.1 --port 8080 -c 4096 -ngl 99

# CLI
llama-cli -m puppygemma-Q4_K_M.gguf -r "<end_of_turn>" -c 4096 -ngl 99

The -r "<end_of_turn>" (reverse prompt / stop) is required — without it the model never emits EOS and loops its answer forever. This is a downstream artifact of the training-time EOS remap and is the one real quirk you'll hit.

Training

Dataset format, generation scripts and full pipeline live in the local project. Retraining with get_chat_template(..., map_eos_token=False) + explicit <eos> in the training text produces a model that stops on its own.

License / provenance

Derived from google/gemma-4-E2B-it — Gemma terms of use apply to the base model weights. The puppy is original research.

Made with love, a 4080 SUPER, and approximately one (1) treat per epoch. 🦴

Downloads last month: 2

GGUF

Model size

5B params

Architecture

gemma4

Hardware compatibility

4-bit

8-bit

Model tree for Xordas/puppygemma

Base model

google/gemma-4-E2B

Finetuned

google/gemma-4-E2B-it

Adapter

(177)

this model