A thousand-upvote Ask HN thread put a question a lot of us have been circling directly to the people who would know: has anyone fully swapped Claude or GPT for a local model as their main coding tool, not just for side experiments? It drew 488 comments. The honest answer from those who've actually done it is more useful than the usual hype, and it points one way: a local model for coding works now, as a junior you have to guide closely, while the people with the best hardware still reach for the cloud on the hard stuff.
The setup everyone cited
The most-quoted rundown came from a commenter running Qwen3.6 35B with only 3B active parameters on a Mac Studio with 128GB RAM (and a 36GB MacBook), fully offline through a sandboxed coding harness. They redesigned a Django + Wagtail site with it, including the lesser-known Wagtail parts the model had no internet access to look up. The line that stuck with me: if Opus gives a 15x speedup, fully offline Qwen gives a 5x speedup, and since it's free, that's still "mind-boggling."
The catch is in the texture. Local models don't think for you. Leave an assumption open and Qwen3.6 takes the easiest route (CSS jammed into HTML), gets into loops, and frequently botches the edit tool call, then burns thinking tokens re-reading files instead of just retrying. The commenter's analogy was a junior with broad knowledge you must guide, versus a senior who thinks with you on architecture. That maps to my own experience with smaller models.
Where local actually breaks
The sharpest test in the thread came from HappySweeney, who runs full-fat models off Optane and RAM at about 0.7 tok/s. Their go-to benchmark: rewrite a scalar bit-matrix transpose using AVX-512. Cloud models "play with that like it's nothing." Kimi 2.6 and GLM 5.1 "both failed miserably." [1]
“I have an optane and lots of ram, so I tried full-fat models for writing some function overnight, as I get about 0.7 t/s. My current go-to test is to update a scalar function to transpose a bit-matrix to one using avx512. the cloud models all play with that like its nothing. Kimi 2.6 and GLM 5.1 both failed miserably.”
That's the real boundary, novel low-level optimization, and it lines up with the ceiling another commenter put on it: Qwen3.6 27B dense is roughly equal to Claude Haiku 4.5, maybe Sonnet depending on the task. [2] Haiku-class is genuinely useful. It is not the frontier, and nobody serious in the thread pretended otherwise.
What builders actually said
Sentiment split cleanly into two camps. The pro-local case rests on privacy and zero cost, and it's real for people who care about it. The skeptical case is about friction and capability. One commenter tried Gemma 4 on an Apple M4 and found the tok/s significantly lower than the cloud offering, adding that the lack of enterprise tooling to pick an appropriate model and run it locally doesn't help. [3]
The most damning data point came from someone with no excuse. arjie runs two RTX Pro 6000 Blackwell cards on DeepSeek V4 Flash at 160 tok/s, with one system auto-writing code and another auto-reviewing it. It's "blazing fast." And yet: "it's mostly habit that keeps me with CC and Codex." [4]
“Not “local” and not interactive coding but sharing since it might be helpful. I have 2x RTX Pro 6000 Blackwell running DeepSeek V4 Flash. I get 160 tok/s raw but it’s a reasoning model. For my use case, I have it auto-write code and another system auto-review the code. I occasionally use it with pi to write some code and it’s blazing fast but it’s mostly habit that keeps me with CC and Codex.”
When someone with that rig still defaults to the cloud, the thing keeping people on Claude and Codex isn't only capability. It's inertia, polish, and trust. The practical-minded reply to all of it came from kertoip_1: stop theorizing, attach OpenRouter to your agent and try the open-weight models yourself, because everyone's needs differ. [5]
My read
The thread doesn't support "local has caught up," and it doesn't support "local is a toy." It supports something more specific. A local model for coding is now a legitimate daily driver for well-scoped, privacy-sensitive work, if you're willing to be precise and to babysit the agent loop. For hard problems, novel algorithms, gnarly refactors, anything where you'd want a senior thinking alongside you, the frontier cloud models still win, and the gap is large enough that even enthusiasts with serious GPUs keep paying for them.
What would change my mind is a single reproducible report of a local setup clearing HappySweeney's AVX-512 bar, or beating Sonnet on a real refactor without hand-holding. Nobody in 488 comments posted one. That absence is the strongest evidence in the thread.
If you're building on this today
Try it, cheaply, before you commit hardware. Point your existing coding agent at OpenRouter's open-weight models and run your own real tasks through them for a week. [5] If a Qwen3.6 35B-class model handles 80% of your work, the privacy and cost win is genuine, keep the cloud subscription for the other 20%.
If you're buying hardware, a 128GB unified-memory Mac or Strix Halo box is the sweet spot people are actually using, and it's worth watching antirez's ds4 project mature for that platform. [6] One commenter floated per-user RLHF to sand off the generic-model tics; interesting, unproven, and probably more effort than it's worth for one person's feedback. [7]
The takeaway I'd bet on: local models are now good enough that the reason to stay on Claude is shrinking from "it's far smarter" toward "it's smarter on the hard 20% and I trust the tooling." That's a different argument than we were having a year ago, and it's moving in one direction.
Sources
Primary source: Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?
- HN comment by HappySweeney — HappySweeney · community
- HN comment by K0balt — K0balt · community
- HN comment by tumetab1 — tumetab1 · community
- HN comment by arjie — arjie · community
- HN comment by kertoip_1 — kertoip_1 · community
- HN comment by christkv — christkv · community
- HN comment by acc_297 — acc_297 · community
- HN comment by dude250711 — dude250711 · community