Intel / post

DR-Venus packs a deep research agent into 4B parameters

Source: x.com
Date: Published Apr 23, 2026

🆕We're open-sourcing DR-Venus — a 4B #deepresearch agent built entirely on open data. Built with two-stage training: SFT with strict cleaning → RL with IGPO turn-level optimization, and Dense credit assignment for 200+ turn trajectories. What's inside: ✅ Training code & recipes ✅ Model checkpoints (GGUF format for edge deployment) ✅Reproducible paper with our technical details #inclusionAI #OpenSource #AIAgent 🤗 Models: https://t.co/XTAqaFEZyt 📄 Paper: https://t.co/S4E31ImXu1 🐙 Code: https://t.co/k9XCyiqwJr #OpenSource #DeepResearch #EdgeAI #AIAgent

Apr 23, 2026 View on X

Context

Training an to do deep, multi-step web research is hard specifically because a single research task can span 200 or more turns, and it's not obvious which of those turns actually deserved credit when the final answer turns out right or wrong. Ant Open Source, under InclusionAI, open-sourced DR-Venus, a 4B-parameter research agent trained in two stages: supervised with strict data cleaning, then reinforcement learning using a method the team calls IGPO for turn-level optimization and dense credit assignment across long trajectories. The release includes training code, GGUF checkpoints sized for local or edge deployment, and a paper describing the method.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.

Read the source x.com