Dev plans to SFT a 4B model into a specialized coding agent, starting with 100-200 mini SaaS apps
ConfidentDinner6648 · reddit · 2026-11-07
A Reddit developer (ccore), who previously released 3D Coder and GLB 3D Coder, is proposing to fine-tune the pretrained Qwen 3.5 4B into a specialized agentic coding model — and is gathering community feedback before training.
Key arguments
- Originally planned SFT followed by DPO/GRPO, but high-quality SFT alone generalized far beyond expectations: the 4B produced complex Three.js scenes, voxel structures and outputs over 64K tokens, including scenes never seen in training.
- Hypothesis: transferring capabilities already demonstrated by stronger models is substantially cheaper than rediscovering them via RL; data quality and consistency matter far more than token count.
- Small models shouldn't try to be good at everything — like a daily car that doesn't need off-road capability, they should excel at everyday tasks.
Experiment plan
- Start with 100–200 functional mini SaaS apps covering development, testing, debugging and maintenance tasks.
- Scale to 100,000 validated multi-turn trajectories across 30–40 subdomains (React/Vue/Next.js, Python/Node.js backends, games and 3D).
- The base 4B already scaffolds Vite projects, installs dependencies and builds simple apps, though inconsistently.
More from coding & agent
- Stardock CEO: orchestrators miss the point—teams need exhibits and shared agent data — draginol · 2026-11-07
- Researcher uses Claude to auto-fix LaTeX figure offsets: "ridiculously painless" — alfcnz · 2026-11-07
- Peer-reviewed paper: symbolic logic + geometric coupling for LLM agent memory belief revision — letandrewcook · 2026-11-07
- Agentic robot control demo: 2.5 hours autonomous assembly at one gadget per 7 minutes — ihorbeaver · 2026-11-07
- Claude Skills at $3: what enterprise teams stopped doing by hand — petrusenko_max · 2026-11-07
- Dials: training-free 14KB steering sliders for MiniMax H3 video in ComfyUI — PurzBeats · 2026-11-07