DevMartin Fowlerabout 2 hours ago
Viability of local models for coding
1 min read
A Thoughtworks engineer tested local LLMs for agentic coding over four weeks on M3 Max and M5 Pro machines. Models between 8GB and 48GB ran on 48GB and 64GB RAM machines. Reasoning sometimes hurt performance. Qwen3 Coder Next 80B solved tasks better but crashed. MoE models offered the best balance.
Level
Hype check
Tap to vote and see what everyone thinks.
