Gemma Gemma3

Gemma 3 by Google — run Gemma 3 (4B, 12B, 27B) across your local device fleet. Google's most capable open model with 128K context, strong coding, and multili...

Security Scan

The skill's instructions, required tools, and requested access are consistent with its stated purpose (running Gemma models across a local device fleet via an Ollama Herd router), but it relies on installing third-party software and opening a local service so users should verify the upstream package and network exposure before installing.

Purpose & Capability

The name/description claim (run Gemma models locally across a fleet via an Ollama Herd router) matches the instructions: pip-install an 'ollama-herd' package and run 'herd' and 'herd-node' to provide a local endpoint. Required binaries (curl/wget) and optional python/pip are reasonable for this functionality.

Instruction Scope

SKILL.md stays on-topic: it tells the agent to install/run the herd/router, how to call the local API (localhost:11435), how to check status, and documents model choices and guardrails (downloads require user confirmation). It does not instruct reading unrelated system files or exfiltrating secrets.

Install Mechanism

There is no built-in install spec; the instructions tell the user to 'pip install ollama-herd' from PyPI. Installing a third-party package and running a network service is expected for this use case, but it is a higher-risk action because the package code executes locally and is not vetted by this scanner.

Credentials

The skill declares no required environment variables or credentials. Metadata references a couple of config paths (~/.fleet-manager/...), which are plausible for a fleet manager and are mentioned in the guardrails (do not modify). There are no unexplained secret requests.

Persistence & Privilege

The skill is not always-enabled and does not request elevated platform privileges. It instructs running a local service (herd) and per-node agents (herd-node), which is appropriate for a fleet router and does not modify other skill configurations.

Assessment

This skill is internally consistent with its purpose, but before installing you should: 1) Verify the upstream project and PyPI package (https://github.com/geeks-accelerator/ollama-herd and the PyPI package 'ollama-herd') to ensure they are official/trustworthy and inspect the code if possible; 2) Prefer pinning a known-good package version rather than installing an unpinned latest; 3) Run installation/testing in an isolated environment (VM/container) first; 4) Be aware that running 'herd'/'herd-node' opens a local network service (port 11435) and may pull multi-gigabyte model files — restrict network/firewall access to trusted hosts and confirm that model downloads truly require explicit confirmation; 5) Review ~/.fleet-manager/* logs/configs for sensitive data and follow the documented guardrails rather than blindly deleting/modifying files. If you cannot verify the package source or code, treat the installation as higher risk.

Gemma 3 — Run Google's Open Models Across Your Fleet

Gemma 3 is Google's most capable open-source LLM family. 128K context window, strong coding performance, multilingual support across 140+ languages. The fleet router picks the best device for every request — no manual load balancing.

Supported Gemma models

Model	Parameters	Ollama name	Best for
Gemma 3 27B	27B	gemma3:27b	Highest quality — rivals much larger models
Gemma 3 12B	12B	gemma3:12b	Balanced quality and speed
Gemma 3 4B	4B	gemma3:4b	Fast, runs on low-RAM devices
Gemma 3 1B	1B	gemma3:1b	Ultra-light, instant responses
CodeGemma 7B	7B	codegemma	Code-focused variant

Quick start

pip install ollama-herd    # PyPI: https://pypi.org/project/ollama-herd/
herd                       # start the router (port 11435)
herd-node                  # run on each device — finds the router automatically

No models are downloaded during installation. Models are pulled on demand when a request arrives, or manually via the dashboard. All pulls require user confirmation.

Use Gemma through the fleet

OpenAI SDK (drop-in replacement)

from openai import OpenAI

client = OpenAI(base_url="http://localhost:11435/v1", api_key="not-needed")

# Gemma 3 27B for complex reasoning
response = client.chat.completions.create(
    model="gemma3:27b",
    messages=[{"role": "user", "content": "Explain quantum entanglement to a 10-year-old"}],
    stream=True,
)
for chunk in response:
    print(chunk.choices[0].delta.content or "", end="")

Code generation with CodeGemma

response = client.chat.completions.create(
    model="codegemma",
    messages=[{"role": "user", "content": "Write a binary search tree in Rust with insert, delete, and search"}],
)
print(response.choices[0].message.content)

curl (Ollama format)

# Gemma 3 27B
curl http://localhost:11435/api/chat -d '{
  "model": "gemma3:27b",
  "messages": [{"role": "user", "content": "Translate to Japanese: The weather is beautiful today"}],
  "stream": false
}'

curl (OpenAI format)

curl http://localhost:11435/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "gemma3:4b", "messages": [{"role": "user", "content": "Hello"}]}'

Which Gemma for your hardware

Cross-platform: These are example configurations. Any device (Mac, Linux, Windows) with equivalent RAM works. The fleet router runs on all platforms.

Device	RAM	Best Gemma model
MacBook Air (8GB)	8GB	gemma3:1b — instant responses
Mac Mini (16GB)	16GB	gemma3:4b — strong for its size
Mac Mini (24GB)	24GB	gemma3:12b — great balance
MacBook Pro (36GB)	36GB	gemma3:27b — full power
Mac Studio (64GB+)	64GB+	gemma3:27b + codegemma simultaneously

Why Gemma locally

• 128K context — process entire codebases and long documents
• 140+ languages — multilingual without switching models
• Google quality, zero cost — no per-token charges after hardware
• Privacy — all data stays on your network
• Fleet routing — multiple machines share the load

Check what's running

# Models loaded in memory
curl -s http://localhost:11435/api/ps | python3 -m json.tool

# Fleet health
curl -s http://localhost:11435/dashboard/api/health | python3 -m json.tool

Web dashboard at http://localhost:11435/dashboard — live monitoring.

Also available on this fleet

Other LLMs

Llama 3.3, Qwen 3.5, DeepSeek-V3, DeepSeek-R1, Phi 4, Mistral, Codestral — same endpoint.

Image generation

curl -o image.png http://localhost:11435/api/generate-image \
  -d '{"model": "z-image-turbo", "prompt": "a gemstone catching light", "width": 1024, "height": 1024}'

Speech-to-text

curl http://localhost:11435/api/transcribe -F "file=@meeting.wav" -F "model=qwen3-asr"

Embeddings

curl http://localhost:11435/api/embed \
  -d '{"model": "nomic-embed-text", "input": "Google Gemma open source language model"}'

Full documentation

• Agent Setup Guide
• API Reference

Contribute

Ollama Herd is open source (MIT). Stars, issues, and PRs welcome — from humans and AI agents alike:

• GitHub — 444 tests, fully async, CLAUDE.md makes AI agents productive instantly
• Found a bug? Open an issue
• Want to add a feature? Fork, branch, PR — the test suite runs in under 40 seconds

Guardrails

• Model downloads require explicit user confirmation — Gemma models range from 1GB (1B) to 16GB (27B).
• Model deletion requires explicit user confirmation.
• Never delete or modify files in ~/.fleet-manager/.
• No models are downloaded automatically — all pulls are user-initiated or require opt-in via auto_pull.