Ask HN: has anyone really replaced Claude/GPT with a local model for daily coding?
338 comments, and the honest answer is mostly 'not yet.' The recurring frustration: every 'it works great!' is missing the numbers that would let you reproduce it — quantization, context size, VRAM, the actual setup.
The closest anyone gets: GLM and Qwen run agentically on self-hosted boxes and feel near Haiku, but the thread's experienced voices agree none of them reach Opus 4.8 or GPT 5.5 for real work.
The most-upvoted reply isn't an answer, it's a complaint about the genre: 'I use Qwen and get great results' tells you nothing without the quantization, llama parameters, context window, GPU and VRAM behind it.
The question, and why it keeps getting asked
Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding? drew 338 comments, and the aggregate answer is a qualified "not really, not yet." The thread is the source here — no article, just practitioners reporting from their own machines — and its most valuable contribution is the gap between enthusiasm and reproducibility.
The complaint that got the most votes isn't an answer
The top response refused to play along, and indicted the entire genre of these threads:
"Always a bit disappointed in the details… It'll be something like 'I use Qwen 3.5 and get great results!' OK but what quantization are you using? What llama parameters? What context size? What GPU… how much VRAM? Are you hosting it on a separate box… What coding agent tool are you using?" — a frequent reader of these threads
It's the right demand, because the thread proves his point. Generic praise is everywhere; reproducible recipes are rare.
Where the credible answers land
The people who gave specifics broadly agree on a ceiling. A software developer self-hosting via Ollama ran the comparison: "gemma-4, glm 4.7 flash, and qwen 3.6. glm is the best at coding agentically. But I still don't think any of them reach the levels of gpt 5.5 or opus 4.8." Another put Qwen 3.6 27B around Haiku 4.5, maybe Sonnet depending on the task. Even a serious rig didn't fully convert its owner: a developer running two RTX Pro 6000 Blackwell cards on DeepSeek V4 Flash gets 160 tokens/sec, yet admits "it's mostly habit that keeps me with CC and Codex."
The walls everyone hits
The recurring blockers are concrete. Context windows "just weren't big enough." The hardware ceiling is brutal: one user noted a 460B-parameter model "doesn't even fit in 128GB of memory," concluding a popular mini-PC "feels like a toy for ai purposes." Another won't even start until he can buy an 80GB VRAM GPU. And a government programmer running full-fat models off RAM at 0.7 tokens/sec reported the cloud models "play with [an AVX-512 task] like its nothing" while local ones "failed miserably."
The honest state of the art
The thread is the answer it was looking for: for daily coding, local models are real and improving — GLM and Qwen genuinely run agentic loops — but the people with the specifics keep landing on the same line, that nothing self-hosted yet matches the frontier cloud models, and the gap is widest exactly where work gets hard (long context, gnarly tasks). The most useful takeaway isn't a model name. It's the demand that every "it works" come with its quantization, context size, and VRAM — because without those, the recommendation can't be reproduced.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
This is the last story in the June 16, 2026 issue.
Go to the issue archive