SynastrIA Networks

Building AI agents people actually want to use — not just another chatbot wrapper.

A solo-driven (for now) brazilian lab focused on fine-tuning, dataset curation, and agentic behavior alignment. Everything we publish here is open-weight and built from scratch — dataset, training pipeline, all of it.

🧬 What we make

OpenFable-4B — Fine-tuned to capture the reasoning style of Claude Fable 5, after that model got geofenced outside the US. The dataset is 100% custom: no public synthetic data, no contamination, just curated examples across coding, math, agentic planning, and security.

Lumen Qwen 3.5 2B — The model behind Lucian, our AI agent platform. Lightweight, tool-calling capable, built to run subagents with actual personality.

🔗 Links

• Website: synastria.dev
• TikTok: @synastria.dev · @openfable

🛠️ Interests

Open-weight LLM research: fine-tuning, dataset curation, behavior alignment, agentic applications. We post the wins and the failures (looking at you, OpenFable V1 CoT leak).