Agentic Coding

The third part of the local-inference series drops from the 35B MoE to the models most people want on a coding workstation: small, dense, 9-billion-parameter agents. I ran two of them, same base and opposite training recipes, through a full agent loop on a real buggy project plus a Copilot-style autocomplete bench. Neither speed nor correctness separates them: the discriminator is agentic path economy, training- and task-shaped. And local can stand in for Copilot warm, with one 16GB asterisk. Continua a leggere