AI Architecture Updates: July 8, 2026

1. Böckeler Examines When Local Models Are Viable for Agentic Coding

Architecture. Writing on Martin Fowler’s site, Thoughtworks’ Birgitta Böckeler examines the practical factors governing whether locally-run LLMs are viable for agentic coding, identifying RAM as the primary constraint, with 15 to 25GB models fitting comfortably on 48 to 64GB machines given 32 to 64K context windows for complex tasks. She finds mixture-of-experts models such as Qwen3.6 35B offer better RAM efficiency than dense models, Q4 quantization preserves acceptable quality while shrinking footprint, and tool-calling reliability is a key failure mode since local models frequently produce malformed structured calls that require self-correction. The piece also notes that reasoning can be counterproductive in smaller local models, which tend to fall into circular thought patterns, making framework and configuration choices as decisive as raw hardware. Source