Running local models on a laptop is now viable for daily coding

Category: AI
Status: discussion
Sources: Vicki Boykis, discussion
Summary: Vicki Boykis published a practitioner write-up on 2026-06-15 reporting that local model inference on a 64GB M2 Mac has become good enough to replace API calls for many tasks. She runs Gemma 4 (gemma-4-26b-a4b, gemma-4-12b-qat), GPT-OSS-20B, Qwen 3 MoE, and Qwen 2.5 Coder through llama.cpp, Ollama, LM Studio, and llamafiles, and reports agentic coding loops running locally at roughly 75 percent of frontier-model accuracy and speed. She names GPT-OSS as the release after which she stopped reaching for hosted models. Remaining limits are slow inference, small context windows tied to RAM, and prompt-template mismatches.
Comments: HN commenters debated which quantizations stay usable on 32GB versus 64GB and reported similar results with MLX on Apple Silicon; several noted that throughput, not quality, is now the main gap.
Why it matters: Capable open-weight models plus mature local tooling shift a class of coding and refactoring work off metered APIs and onto developer hardware.