Boykis argues that with a mundane setup — an M-series MacBook, Ollama/LM Studio, and open-weights models like Qwen, Llama 3.x, and Gemma — local LLMs are now sufficient for the median knowledge-work task. She is not claiming parity with frontier models, but that the gap has narrowed enough that drafting emails, summarizing PDFs, and rubber-ducking code work fine locally.
The editorial frames Boykis's narrow claim as an implicit concession: model quality saturated around the GPT-4 threshold for tasks like summarization, rewriting, and explanation, so a 32B Qwen on an M3 Max is sufficient. But agentic coding and other frontier-demanding tasks remain a cloud-model domain, and the two camps in the HN thread are talking past each other because they're describing different workloads.
A recurring objection in the thread is that local models remain inadequate for agentic coding — multi-step tool use, large-context refactors, and autonomous code generation. These commenters concede local models may be fine for prose tasks but argue the frontier cloud models are still required for serious development work.
Many replies converged on the same setup: MLX on Mac or llama.cpp on Linux, with Qwen 2.5 32B as the preferred model. This suggests the practitioner community has settled on a small set of tools and weights that reliably deliver usable performance on consumer hardware.
Vicki Boykis — ML engineer, author of *What Are Embeddings*, and one of the more sober voices in applied ML — published a post titled *Running local models is good now* that landed at 337 points on Hacker News within a day. The piece is short, declarative, and the argument is in the title: the era of 'local models are a fun experiment but you'd never actually use one' is over.
Her setup is mundane on purpose. An M-series MacBook. Ollama and LM Studio for serving. Qwen, Llama 3.x, and Gemma weights pulled from Hugging Face. No exotic quantization, no GPU rental, no Modal endpoint. She uses these models the way most developers actually use LLMs day-to-day: drafting emails, summarizing PDFs, asking dumb questions about unfamiliar code, rubber-ducking architecture decisions. Her claim is not that local models match GPT-5 or Claude 4.6 — it is that the gap between 'frontier' and 'local' has finally narrowed enough that, for the median knowledge-work task, the local model is fine.
The HN thread quickly bifurcated along predictable lines. Half the replies were people sharing their own local stacks (mostly MLX on Mac, llama.cpp on Linux, Qwen 2.5 32B as the consensus pick). The other half were the same comment we've been seeing for two years: *but it can't do agentic coding*. Both groups are correct. They are also talking past each other.
The interesting thing about Boykis's post is what it implicitly concedes by being so narrow. She is not claiming the local model has caught up. She is claiming the workload has split.
For a large class of tasks — summarize this 40-page PDF, explain this stack trace, rewrite this paragraph in a less corporate tone, turn these bullet points into a draft — model quality saturated somewhere around the GPT-4-class threshold. Past that point, the marginal cloud token doesn't make your email better. A 32B Qwen running at 30 tokens/sec on an M3 Max is sufficient. The cloud advantage on these tasks is now mostly latency and convenience, not capability.
For a different class of tasks — multi-step agentic coding, long-context refactors across a repo, tool-use chains with 50+ turns, anything requiring a 200k+ context window held coherently — the gap is still enormous and arguably *widening*. Claude Code, Cursor's background agents, and the Codex CLI we ship on top10.dev's own EC2 box all depend on frontier capabilities that no local quant currently delivers. Yesterday's Ask HN ('Has anyone fully swapped to a local model for coding?') had the same answer it had six months ago and twelve months ago: no, not really, the workflow falls apart by turn three.
This bifurcation is the practitioner insight that gets lost when the discourse collapses into 'is local good yet'. Local models are good for the workload most people *say* they use LLMs for. Local models are not good for the workload most developers *actually* use LLMs for when they're being honest. Both can be true. And Boykis's post is correctly aimed at the first group, which is much larger than HN comment volume suggests.
There's also a quieter argument running under the surface, which is about cost and ownership. Once your daily LLM use exceeds about two hours of active interaction, the math on a $20–60/mo Pro subscription stops being obviously better than amortizing an M-series Mac you were going to buy anyway. Add in the privacy story for anyone working on regulated data, the offline story for anyone who travels, and the no-rate-limits story for anyone who has watched their Cursor or Claude budget evaporate at 2pm on a Tuesday — and the local stack starts looking like a real second tool in the toolbox, not a hobby project.
The practical takeaway is to stop running the wrong benchmark. If you tried Ollama in 2024, asked it to write a CRUD endpoint, watched it hallucinate an import, and concluded 'local isn't ready', you ran the test that local was guaranteed to fail. Try it on the workloads it can actually win.
A reasonable 2026 split looks like this: keep the frontier API for code generation, agentic flows, and any reasoning task where you'd notice a quality regression. Move the long tail — meeting note summarization, email triage, doc Q&A over your own files, throwaway research questions, anything you don't want logged on someone else's server — to a local 30B-class model. Qwen 2.5 32B Instruct and Llama 3.3 70B (heavily quantized) are the current consensus picks on 36GB+ Apple Silicon. On a 16GB machine, Gemma 3 12B or Qwen 2.5 14B are the realistic ceiling and they are genuinely usable for the workloads above.
The second-order effect, which Boykis doesn't explicitly call out, is on the data layer. Once a meaningful chunk of your LLM calls runs locally, you stop building infrastructure around 'every prompt is a billable API call'. You can be wasteful. You can pre-process. You can run a model over every file in a repo to build an embedding index without watching a meter tick. That changes the shape of what's worth building.
The interesting question for the next twelve months isn't whether local will catch up to frontier on coding — by every available signal, it won't. The interesting question is whether the *non-coding* category continues to expand. If GPT-5-class quality on summarization and drafting is now achievable in a 32B local quant, it's reasonable to expect that GPT-5-class quality on tool use, structured output, and short-horizon agents will follow in 2027. The frontier doesn't have to stop moving for local to keep eating the workloads behind it. Boykis's post is worth taking seriously not because local has won, but because the line between 'use the API' and 'use the laptop' has finally moved from a religious argument into a workload-dependent engineering decision. That is the kind of progress that doesn't trend on Hacker News for long, which is usually a sign it's real.
After having been a happy user of Qwen3.6-27B for a few weeks, due to being away from the hardware, I'm currently forced to use Claude Sonnet 4.6It is such a downgrade. I don't understand how that's even possible. The thing has so many strongly-held opinions I did not ever ask it for,
This is the kind of thing that Anthropic et al should be worried about. As it becomes easier and easier to run local models, the ceiling of what they'll be able to charge will get lower and lower. Not that nobody will be willing to pay $$$$$ per month, but a lot of people are going to multiply
[meta] I wonder why people have such wildly different bar for what is "good" agentic coding?In a way, it's absolutely amazing that we've went from "Playing 'Set a Timer' on Apple Music" intelligence to something that may pass the Turing Test, but in practical
Show us the resulting code of using them! :) I want to use local models, I have the hardware for it, but while trying them out as replacements for GPT 5.5 xhigh or Opus or other SOTA models, they aren't quite ready to be replaced yet, sadly. The quality and bumps they encounter just slows down
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I don't know about good, I use a lot of local models and they're still pretty painful to run locallyYou have dense models (qwen 27b, gemma 31b) who are pretty smart, but pretty slowYou have MoE models (gemma 26b, qwen 35b, north mini code 30b) who are pretty fast, but make a lot of mistake