Independent Developer Plans to Train 4B Model as Lightweight Coding Agent

On November 7, 2026, an independent developer posted on the LocalLLaMA subreddit on Reddit to lay out plans for training a small 4B model into a capable coding Agent built to compete with much larger models. The developer explained that they had previously shipped a series of 3D coding models, including 3D Coder and GLB 3D Coder, and drew a lesson from that process: the community should have been brought in earlier. According to the post, community feedback directly shaped the second model and could have saved the developer a number of additional training runs. For this project, the developer wants to discuss the idea with the community before any training begins, with the goal of gathering feedback to refine the training roadmap. Explaining why the approach is feasible, the developer noted that when they first trained 3D Coder, the original plan was to use full supervised fine-tuning (SFT) to build baseline capabilities, followed by DPO or GRPO to further improve the model's performance. At the time of writing, the plan remains at the discussion stage, with training yet to start.