On Hacker News, users debated whether anyone has fully swapped cloud‑based assistants like Claude or GPT for a locally hosted model as a primary coding companion. The thread asks for concrete setups, token‑per‑second rates, and performance, not just hobby experiments. Participants note that fine‑tuning on personal prompts could prune unwanted verbosity or sycophancy, but the effort may introduce unpredictable behavior and loss of proprietary pipelines.

Several contributors listed their rigs. One runs Ollama with Gemma‑4 and GLM‑4.7‑Flash on an Optane‑rich server, achieving roughly 0.7 tokens / s on overnight function generation for nightly builds. Another uses two RTX 6000 Blackwell cards to host DeepSeek V4 Flash, reaching 160 tok/s raw on reasoning tasks and pairing it with an auto‑review pipeline. A third reports Qwen 3.6 27B dense delivering code quality comparable to Claude Haiku 4.5.

Respondents repeatedly cite memory bandwidth and VRAM caps as blockers; a 460 billion‑parameter model exceeds 128 GB even on high‑end GPUs, forcing many to settle for smaller variants. Sparse enterprise tooling for model selection, quantization, and integration with IDEs further hampers adoption. The thread illustrates that, while functional local coding agents exist, reproducible, high‑throughput setups remain scarce.