Download
Gemma 4
Open Models

from transformers import AutoProcessor, AutoModelForImageTextToText
import torch

# Load model — swap model_id for your chosen size
model_id = "google/gemma-4-26b-it"   # or e2b-it / e4b-it / 31b-it

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

messages = [{
    "role": "user",
    "content": [{"type": "text", "text": "Explain quantum entanglement."}]
}]

inputs = processor.apply_chat_template(
    messages, tokenize=True,
    return_dict=True, return_tensors="pt",
    add_generation_prompt=True,
    enable_thinking=True,           # ← thinking mode!
).to(model.device)

output = model.generate(**inputs, max_new_tokens=2048)
print(processor.decode(output[0], skip_special_tokens=True))

# pip install google-genai
from google import genai

client = genai.Client(api_key="YOUR_API_KEY")

response = client.models.generate_content(
    model="gemma-4-31b-it",       # or gemma-4-26b-it
    contents="Write a haiku about open-source AI.",
)
print(response.text)

# ── REST equivalent ─────────────────────────────────
# curl "https://generativelanguage.googleapis.com/v1beta/
#   models/gemma-4-31b-it:generateContent?key=YOUR_KEY" \
#   -H 'Content-Type: application/json' -X POST \
#   -d '{"contents":[{"parts":[{"text":"Hello!"}]}]}'

# Install vLLM
pip install vllm

# Start OpenAI-compatible server
vllm serve google/gemma-4-26b-it \
  --dtype auto \
  --max-model-len 65536 \
  --tensor-parallel-size 2    # for multi-GPU

# Query via OpenAI SDK
from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="none",
)

resp = client.chat.completions.create(
    model="google/gemma-4-26b-it",
    messages=[{"role":"user","content":"Hello Gemma!"}],
)
print(resp.choices[0].message.content)

# Install MLX LM
pip install mlx-lm

# Run inference (4-bit quantized)
mlx_lm.generate \
  --model mlx-community/gemma-4-e4b-it-4bit \
  --prompt "Explain transformers simply." \
  --max-tokens 512

# Or 26B at 4-bit for M2 Max / M3 Max (≥64 GB)
mlx_lm.generate \
  --model mlx-community/gemma-4-26b-it-4bit \
  --prompt "Write a Python quicksort." \
  --max-tokens 1024

# Python API
from mlx_lm import load, generate

model, tokenizer = load("mlx-community/gemma-4-e4b-it-4bit")
response = generate(model, tokenizer, prompt="Hello!", max_tokens=256)
print(response)

# NVIDIA NIM — containerised, production-ready
docker run --gpus all \
  -p 8000:8000 \
  nvcr.io/nim/google/gemma-4-27b-it:latest

# Unsloth Fine-tune Studio (Mac / Linux / WSL)
curl -fsSL https://unsloth.ai/install.sh | sh
unsloth studio -H 0.0.0.0 -p 8888

# Dockerfile snippet for custom build
FROM nvidia/cuda:12.4.0-runtime-ubuntu22.04
RUN pip install vllm transformers
COPY serve.py /app/serve.py
CMD ["python", "/app/serve.py"]