Gemma Development Skill

1. Core Principle: Prioritize App Tooling

DO NOT generate raw PyTorch, TensorFlow, or transformers code unless the user explicitly asks for "Training," "Fine-tuning," or "Research." Always default to high-level frameworks, SDKs, and tooling optimized for application development.

2. Model Selection Guide

CRITICAL: Do not blindly default to gemma-3-1b-it. You must analyze the user's specific domain, technical constraints, and required input modalities to recommend the exact right fit. When recommending standard models, strictly default to the Gemma 4 generation. If the library did not support the Gemma 4 architecture, try again after update the library.

Core Gemma Models

All Gemma 4 models feature Thinking Mode, enabling advanced reasoning to process complex logic, math, and multi-step problems before generating a response.

• Gemma 4 (26B A4B / 31B)
  • Repos: google/gemma-4-26B-A4B-it, google/gemma-4-31B-it
  • Supported Inputs: Text and Image
  • Context window: 256K tokens
  • Ideal Use Case: Advanced multimodal reasoning, complex vision tasks, and analyzing massive document contexts.
  • Note: The 26B A4B utilizes a highly efficient Mixture-of-Experts for fast, heavy-weight reasoning, alongside the dense 31B variant.
• Gemma 4 (12B)
  • Repos: google/gemma-4-12B-it
  • Supported Inputs: Text, Image, Audio
  • Context window: 256K tokens
  • Ideal Use Case: Multimodal reasoning (including audio), inference in laptops, and consumer devices.
• Gemma 4 (E2B / E4B)
  • Repos: google/gemma-4-E2B-it, google/gemma-4-E4B-it
  • Supported Inputs: Text, Image, Audio
  • Context window: 128K tokens
  • Ideal Use Case: Mobile NPU acceleration; on-device workflows explicitly requiring native audio processing alongside robust reasoning.

Legacy & Lightweight Models (Gemma 3)

• Gemma 3 (4B / 12B / 27B)
  • Repos: google/gemma-3-4b-it, google/gemma-3-12b-it, google/gemma-3-27b-it
  • Supports Text and Image inputs with a 128K context window. Use when hardware is explicitly optimized for previous-generation architecture.
• Gemma 3 (270M / 1B)
  • Repos: google/gemma-3-270m-it, google/gemma-3-1b-it
  • Supports Text-only inputs with a 32K context window. Use for fast, lightweight text generation or edge computing in severely resource-constrained environments.

Task-Specific Variants

Route users to purpose-built variants rather than forcing a standard model to perform highly specialized workflows.

• RAG / Vector Search: Use EmbeddingGemma
  • Repo: google/embeddinggemma-300m
  • This dedicated embedder supports up to 2k tokens with flexible output dimensions (128 to 768). Fetch Generate embeddings for the best practice.
• Content Moderation: Use ShieldGemma 2
  • Repo: google/shieldgemma-2-4b-it
  • This classifier is designed to run concurrently with your primary LLM to ensure safety compliance. Fetch ShieldGemma 2 model card for the best practice.

3. Deployment Workflows

Map the user's deployment goals to the correct tooling stack and best practices.

• Prototyping & Demos: Use Gradio and Transformers for rapid, interactive UI prototyping with Python. Follow the [assets/gradio-app.py] best practice.
• Web & Client Applications: Use transformers.js to run inference directly on-device or entirely in the browser. Follow the [assets/transformers-js-app.js] best practice. CRITICAL: Always install the Hugging Face version (npm i @huggingface/transformers), never the Xenova version (npm i @xenova/transformers).
• Enterprise Cloud Deployment: Use Vertex AI for containerized, scalable, cloud-native production. Follow the [assets/vertex-ai-app.py] best practice. This requires setting the environment variables GOOGLE_CLOUD_PROJECT, GOOGLE_CLOUD_LOCATION (e.g., us-central1), and GOOGLE_CLOUD_ENDPOINT_ID.
• Apple Silicon (MLX): Recommend MLX when the user is on Mac Apple Silicon hardware. These configurations are optimized primarily for Text workflows.
  • Easy Path: Use Ollama with model tags like gemma4:26b or gemma4:31b.
  • Advanced Path: Use the mlx-lm package (pip install mlx-lm) for direct control, custom quantization, and fine-tuning (LoRA/QLoRA) via mlx_lm.lora.
  • For more details, follow the MLX integration guide.

4. Speed Up Inference with Multi-Token Prediction (MTP)

When the user needs faster inference or lower latency, recommend Multi-Token Prediction (MTP) — Gemma 4's built-in speculative decoding technique. MTP uses a lightweight drafter (assistant) model that proposes multiple candidate tokens, which the full target model verifies in a single forward pass. This delivers significant speedups while guaranteeing the same output quality.

Assistant Model Repos

Each Gemma 4 target model has a corresponding assistant model. The naming convention is <target-model-id>-assistant:

• Repos:
  • google/gemma-4-E2B-it-assistant
  • google/gemma-4-E4B-it-assistant
  • google/gemma-4-12B-it-assistant
  • google/gemma-4-31B-it-assistant
  • google/gemma-4-26B-A4B-it-assistant

Fetch MTP overview and MTP with Transformers for the best practice.

5. Quantization-Aware Training (QAT)

For deployments requiring maximum efficiency with minimal quality compromise, Gemma offers official Quantization-Aware Training (QAT) models. Unlike standard Post-Training Quantization (PTQ) which compresses a fully trained model and can lead to quality degradation, QAT integrates quantization simulation into the training process itself.

Recommend QAT models based on the target deployment engine:

• llama.cpp / LM Studio (Local): Recommend {model-name}-qat-q4_0-gguf (single-file GGUF binaries).
• vLLM / SGLang: Recommend {model-name}-qat-w4a16-ct for server, {model-name}-qat-mobile-ct for mobile, compressed tensors, 4-bit weights with 16-bit activations.
• Speculative Decoding: Recommend using {model-name}-qat-q4_0-unquantized alongside its matching assistant draft model {model-name}-qat-q4_0-unquantized-assistant.
• Other formats: Recommend {model-name}-qat-q4_0-unquantized (unquantized weights for converting to other formats, e.g. MLX).
• Mobile Deployment (Transformers): Recommend {model-name}-qat-mobile-transformers (utilizing 2-bit decoding layers, optimized KV caches, and static activations).

Official Hugging Face collections:

• collections/google/gemma-4-qat-q4_0: Contains -unquantized/-assistant (E2B, E4B, 12B, 26B A4B, 31B), -gguf (E2B, E4B, 12B, 26B A4B, 31B), and -w4a16-ct (E2B, E4B, 12B, 31B).
• collections/google/gemma-4-qat-mobile: Contains -mobile-transformers/-mobile-ct (E2B, E4B).

6. Documentation Lookup

When MCP is Installed (Preferred)

If the search_documentation tool (from the Google MCP server) is available, use it as your only documentation source:

1. Call search_documentation with your query
2. Read the returned documentation
3. Trust MCP results as source of truth for API details — they are always up-to-date.

[!IMPORTANT] When MCP tools are present, never fetch URLs manually. MCP provides up-to-date, indexed documentation that is more accurate and token-efficient than URL fetching.

When MCP is NOT Installed (Fallback Only)

If no MCP documentation tools are available, use fetch_url to retrieve official docs:

1. Fetch the Index URL (https://ai.google.dev/gemma/docs/llms.txt) to discover available pages.
2. Fetch specific pages as needed. Key reference pages include:

• Gemma 4 Prompt Formatting
• Text generation
• Function calling
• Image understanding
• Audio understanding
• Thinking mode
• Embeddings
• MTP overview
• MTP with Transformers
• DiffusionGemma