gemma-dev

What the source does

• FilesystemNone observedNo reads or writes in the source
• NetworkNone observedNo outbound calls in the source
• ShellNone observedNo commands run in the source
• DependenciesPinnedNothing floats
• SecretsNone foundNo keys in source

Five audit layers

Gates applied: no behavioural pass.

Audit 0.4.1 · source sha 524f4804e9cd · 20 Oct 2026Report an issue or request a re-scan

• Publishergoogle-gemma GitHub organizationOwner of the repository this skill is read from
• Source foldergoogle-gemma/gemma-skills/skills/gemma-devThe exact folder we audited
• LicenceApache-2.0As the repository states it
• Commit pinned for this verdict524f4804e9cdA new audit runs when the source changes
• Repository statusNot archivedLast push 2 days ago
• Sandboxed runNot runSkills are read, not executed; layer L4 is skipped

Its description: Trigger this skill when building applications with Gemma or for general knowledge inquiries related to Gemma models (e.g. prompt structure, capabilities). Covers model selection, development workflows, and deployment…

---
name: gemma-dev
description: Trigger this skill when building applications with Gemma or for general knowledge inquiries related to Gemma models (e.g. prompt structure, capabilities). Covers model selection, development workflows, and deployment best practices.
---

# 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 2**
  - Repo: `google/embeddinggemma-2`
  - This dedicated embedder supports up to 8k tokens with flexible output dimensions (128 to 768). Fetch [EmbeddingGemma 2 model card](https://gemma.ai/models/embeddinggemma/model-card-v2.md.txt) 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](https://gemma.ai/models/shieldgemma/model-card-v2.md.txt) 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](https://ollama.com/library/gemma4/tags) 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](https://gemma.ai/deployment/mlx.md.txt).

## 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 [Speed-up Gemma 4 with Multi-Token Prediction](https://gemma.ai/get-started/capabilities/mtp.md.txt) 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 comp

Excerpt, first 6,000 characters of SKILL.md, verbatim from the audited commit. It is text from someone else’s repository, quoted so you can judge it before your agent reads it.

git clone https://github.com/google-gemma/gemma-skills
git -C gemma-skills checkout 524f4804e9cd

Works in

1. 2026-10-20
   ✓SAFE B 89 / 100

   first audit · source 524f4804e9cd

Questions about gemma-dev

• github.com/google-gemma/gemma-skills/tree/main/skills/gemma-dev — the audited source at commit 524f4804e9cd and every finding aboveObserved · 20 Oct 2026

Spotted something wrong? Report a correction · Methodology