Introducing GreenPT Code: coding models built for agents

We just launched GreenPT Code: a suite of specialized open-weight AI models for software development, served from one OpenAI-compatible API on sustainable European infrastructure.

GreenPT Code gives coding agents five open-weight models, from glm-5.2 with its 1M-token context down to an efficient 30B Mixture-of-Experts coder, all hosted on renewable EU energy, with no training on your code and CO2 reported per request.

Here’s why we built it.

Coding agents deserve better defaults

AI coding assistants have quietly become the heaviest AI workload most teams run. An agent working through a refactor doesn’t send one prompt, it sends hundreds: planning steps, tool calls, file edits, test runs. Every one of those requests lands on a GPU somewhere.

For most teams, “somewhere” means a proprietary model in a US data centre, trained on who knows what, with your codebase as potential training data.

We think you can have the capability without those trade-offs. The best open-weight coding models have caught up: they plan across steps, call tools reliably, and edit many files at once. What was missing was a place to run them that takes privacy and energy seriously.

Five open models, one API

We do not train secret models. Every GreenPT Code model is open-weight, and we host them ourselves in green EU data centres:

Model	Builder	Context	Best for
glm-5.2	z.ai	1M	Our flagship. Long-horizon agentic engineering
kimi-k2.7-code	Moonshot AI	256K	Codebase-scale reasoning, 1T MoE
devstral-2-123b-instruct-2512	Mistral	256K	SWE agents, multi-file editing
qwen3-coder-30b-a3b-instruct	Qwen (Alibaba)	256K	Efficiency: 3.3B of 30.5B params active
minimax-m2.5	MiniMax	200K	Agentic coding on a budget

Start with glm-5.2. It’s the recommended default for AI coding assistants: reliable tool calling, multi-step planning, and multi-file refactoring over large codebases, with a context window big enough to reason over an entire repository in one pass.

The Mixture-of-Experts designs matter for sustainability too. qwen3-coder activates only 3.3B of its 30.5B parameters per request, so it does the work of a much larger dense model at a fraction of the compute and energy.

Connect your agent in three values

GreenPT Code is OpenAI-compatible, so any tool that supports a custom endpoint works unchanged:

Base URL:  https://api.greenpt.ai/v1
API key:   GREENPT_API_KEY
Model:     glm-5.2

We wrote step-by-step guides for OpenCode, Cline, Aider, Kilo Code, Claude Code, and Codex CLI. Cursor, GitHub Copilot BYOK, Continue, and similar tools connect with the same three values.

Sustainable, measured, not claimed

The models run in Scaleway’s green EU data centres with a PUE of 1.25 and a WUE of 0.25, well below industry averages. And because “green” claims are easy to make, we report CO2 and energy usage per API call, so you can see exactly what your agent consumes.

Privacy gets the same treatment as sustainability: your code is processed entirely inside the EU, with full GDPR compliance and data processing agreements on request. We never train models on your code or conversations. Not opt-out. Never.

Cut what your agent emits

Hosting efficient models is half the equation. The other half is making agents emit less in the first place, because every token your agent doesn’t generate is cost, latency, and energy saved.

That’s why GreenPT Code pairs with our open-source agent skills:

• Honey trims code and prose together: around 49% fewer code tokens at 98% of baseline quality.
• Ponytail writes minimal, YAGNI-first code: around 54% less code for the same task.
• Caveman keeps the work and drops the words: around 75% fewer output tokens.

Efficient models, running on renewable energy, generating fewer tokens. The savings compound.

Try it

Start a free 14-day trial, no credit card needed. Create an account, point your coding agent at the endpoint above, and you’re running open models on 100% renewable energy.

The full model specs and setup guides live in the coding docs.