Google released Gemma 4 on 2 April 2026 under the Apache 2.0 licence — and within weeks, practitioners across X were building fully local coding agents with it. The combination of Gemma 4's native function-calling, Ollama's OpenAI-compatible API, and Apple Silicon's MLX backend has made privacy-first AI coding assistance genuinely practical for small and mid-sized businesses.
Developer and AI educator Patrick Loeber (@patloeber) describes his setup as "running coding agents fully locally" using Gemma 4 26B, the Pi agent framework, and either LM Studio or Ollama as the inference server. The entire stack runs on local hardware — no cloud connection, no token bills.
What Is Gemma 4?
Gemma 4 is Google's open-weight model family, available in four variants sized for different hardware budgets:
| Variant | Total params | Active params | VRAM (Q4) | Target hardware |
|---|---|---|---|---|
| E2B | 2 B | 2 B | ~5 GB | Laptop, mobile edge |
| E4B | 4 B | 4 B | ~5 GB | Laptop, dev workstation |
| 26B MoE | 26 B | ~3.8 B | ~16 GB | GPU workstation, Mac |
| 31B Dense | 31 B | 31 B | ~24 GB (Q4) | Mac Studio, server |
The standout is the 26B Mixture-of-Experts (MoE) variant: it activates only 3.8 billion of its 26 billion parameters per inference pass. That delivers generation speed close to a 4B model while the model draws on the knowledge encoded across the full 26B parameter set — an efficient trade-off for interactive coding assistance.
All variants ship with native function calling (required for agent tool use), multimodal input (text and images), and a 128K-token context window that accommodates sizeable code repositories within a single session.
Local Coding Agents: What Practitioners Are Building
Several production-ready setups have emerged since April 2026:
Pi agent + Gemma 4 26B
Pi agent connects to Ollama's OpenAI-compatible endpoint (http://localhost:11434/v1) and uses Gemma 4 as its reasoning engine. The agent reads and writes files, executes shell commands, and works through multi-step programming tasks autonomously — entirely offline. Source code, error logs, and commit history never leave the local machine.
OpenClaw + Gemma 4
OpenClaw, an open-source agent framework with over 250,000 GitHub stars as reported by the community, integrates with Ollama in a setup that takes under ten minutes according to the project's published documentation. The result is a fully local coding assistant with file access and command execution, running at zero marginal cost after the initial hardware investment.
Android Studio — official endorsement
Google added official Gemma 4 support in Android Studio for agentic coding, as documented in the Android Developers Blog. An IDE maker endorsing a local open-weight model for agent workflows is a strong signal that local AI coding assistance has crossed from experiment into mainstream tooling.
Claude Code and OpenCode on Apple Silicon
Ollama notes on X that its MLX-powered Apple Silicon backend makes the update relevant for coding agents that use Ollama as a local server — including Claude Code and OpenCode. The updated stack "unlocks much faster performance to accelerate demanding work on macOS."
Hardware Requirements and Reported Performance
Three practical tiers cover most SMB scenarios:
Entry tier: Gaming laptop or workstation (8–12 GB VRAM) Gemma 4 E4B in Q4KM quantisation runs on most modern laptops with a discrete GPU. As reported by the community, expect 15–25 tok/s on GPU-assisted inference — workable for interactive chat, slower for automated batch code generation.
Mid tier: NVIDIA RTX 3090 / RTX 4080 (16–24 GB VRAM) The 26B MoE model reaches 35–45 tok/s on an RTX 3090 at Q4 quantisation according to reported benchmarks — comparable to a responsive cloud assistant, with no network latency and no per-token cost after initial hardware outlay.
High tier: Mac Studio M3 Ultra (192–512 GB Unified Memory) Ollama's MLX backend for Apple Silicon delivers a further 15–25 % throughput improvement on Mac Studio hardware according to community reports. The 31B Dense model fits entirely in unified memory on a fully configured M3 Ultra. For teams running multiple models in parallel, the Mac Studio's memory architecture is a practical advantage — more on this at Freshlab's local AI page.
All speed figures come from community-reported benchmarks and vary with quantisation level, context length, and specific hardware configuration.
GDPR and Data Sovereignty
For European SMBs, the practical GDPR argument for local AI coding assistance is straightforward.
Cloud-based coding assistants send — depending on provider and settings — code snippets, error messages, comments, and context files to external servers. When source code contains proprietary business logic, customer identifiers embedded in migration scripts, or security-critical configuration values, that creates real data protection exposure. Based on our reading of the current regulatory framework, using an external processor for code would require a Data Processing Agreement, an entry in the Record of Processing Activities, and potentially a Transfer Impact Assessment if the provider processes data outside the EEA.
Gemma 4 running locally via Ollama has no external network connection. That is an architectural property, not a privacy setting or policy commitment. Nothing leaves the local network.
This structural approach to data containment significantly simplifies GDPR compliance documentation and closes the most common gap: sensitive data being processed on infrastructure the business does not control and cannot audit. For a fuller treatment, see our guide on data sovereignty with local AI.
Total Cost of Ownership
A five-person development team with active AI assistant use generates roughly 1–5 million tokens per day under typical usage patterns. At mainstream cloud coding-assistant pricing, that translates to an estimated €30–80 per user per month at intensive use — costs that compound without generating any owned infrastructure.
A local stack built around a used RTX 3090 (market price approximately €600–900 at time of writing) amortises against those ongoing costs within an estimated 12–18 months at typical team usage intensity. Beyond that point, the only operating cost is electricity: a RTX 3090 draws around 350 W under load, adding roughly €25/month at €0.30/kWh with eight hours of active daily use.
Apple Silicon hardware carries even lower operating costs per token due to its superior performance-per-watt profile.
Getting Started
# Install Ollama (macOS/Linux)
curl -fsSL https://ollama.com/install.sh | sh
# Pull Gemma 4 26B MoE (~17 GB download)
ollama pull gemma4:26b
# Or the compact E4B for 8 GB VRAM devices
ollama pull gemma4:4b
# OpenAI-compatible endpoint for any coding agent
# http://localhost:11434/v1Any coding agent that supports an OpenAI-compatible API — Pi agent, OpenClaw, the VS Code Continue extension, or Freshlab's Kaira Toolkit — connects directly to this endpoint. No API key, no internet connection required.
Business Case Summary
Gemma 4 resets the local-vs-cloud calculus for AI coding assistance. The 26B MoE variant delivers competitive inference speed on hardware many SMBs already own or can acquire inexpensively, under an open licence, with no recurring token costs and no data leaving the company network.
For European software teams, the GDPR angle strengthens the case further: architectural data containment beats contractual promises, and the documentation overhead drops substantially when no external processor is involved.
If you want to evaluate a local AI coding stack for your team without committing to infrastructure upfront, Freshlab runs structured pilot projects that take you from model selection through integration and GDPR documentation. Start the conversation today.