The Signal #026 — Dakota’s read on the AI news that actually matters to people running a business.

A thread appeared on Hacker News this week with nearly 1,200 upvotes and close to 500 comments. The question was simple: has anyone fully replaced Claude or GPT with a local model for daily coding work? Not as an experiment. As the main tool.

The answers were surprisingly detailed. And they carry a signal worth reading if you are responsible for any kind of AI tooling decision.

What happened

In a thread posted to Hacker News, a user named cloudking asked whether anyone had genuinely swapped frontier cloud models (AI tools like Claude or GPT that run on remote servers owned by the AI company) for local models (AI that runs entirely on your own hardware, with no data leaving your machine).

One of the top responses came from a user called Greenpants, who shared a detailed breakdown. Running a model called Qwen3.6 35b on a Mac Studio with 128GB of RAM, they described the experience as roughly a 5x speedup for coding tasks. For comparison, they put Claude Opus at roughly a 15x speedup. Their conclusion: local and fully offline gives you 5x, Claude gives you 15x, and the local option costs nothing.

The same user noted a meaningful difference in how the two experiences feel. Their words: local agentic AI is “like a junior with knowledge across the board, that you really need to guide, versus a senior that thinks with you on architecture.”

Another commenter described running a similar setup on a laptop with 128 gigabytes of unified memory, reaching for Qwen 3.6 35B most often for coding and keeping several other models around for different tasks including translation and audio work. Technical discussion in the thread covered things like prompt caching (a technique that lets the model reuse previous conversation context instead of reprocessing it from scratch each time, which affects both speed and cost) and specific configuration flags that improve performance.

The thread is technical. But the underlying question is not.

Why it matters for operators

Most operators are not running local models today, and most probably should not be. The hardware requirements alone, machines with 36 to 128 gigabytes of RAM, put this out of reach for casual use. And as one commenter noted plainly, the performance gap is real. A 5x productivity gain versus a 15x gain is not a rounding error.

But here is what is worth paying attention to.

The people in this thread are not hobbyists. They are practitioners making deliberate decisions about their toolchains. When developers at this level start running serious comparisons and publishing real numbers, it usually means the gap between local and cloud is closing faster than the headlines suggest.

There are also two reasons operators reach for local models that have nothing to do with cost. The first is data privacy. A local model processes everything on your hardware. Nothing is sent to an external server. For a law firm, a medical practice, or any business handling sensitive client information, that distinction is not minor. It may be the whole point.

The second reason is predictable pricing. Cloud models bill per token (a small chunk of words the AI reads or processes). When usage scales up, so does the bill. A local model has a fixed hardware cost and no variable API fees. For high-volume use cases, that math changes the conversation.

What most people get wrong

The framing of “local versus cloud” tends to get treated as a permanent either-or choice. It is not.

The developers in this thread are not abandoning cloud models entirely. Several are running local models for routine tasks and reaching for frontier models when complexity demands it. One user keeps a lineup of different models for different jobs: one for coding, one for translation, one for audio, and others on standby for testing.

That is the actual pattern worth noticing. Operators who get the most out of AI tooling tend to stop asking which single tool is best and start asking which tool is right for which task. A customer service team at an e-commerce company might use a cloud model for nuanced escalations and a local or smaller model for routine classification and tagging. A SaaS company might run expensive frontier models for sales-facing outputs and cheaper or local models for internal document processing.

The tradeoff is not cloud versus local. It is cost, capability, and data sensitivity weighed against each other for each specific job.

The takeaway

This Hacker News thread is not a case study in switching away from Claude or GPT. It is a case study in what it looks like when experienced practitioners start doing real cost-benefit analysis on their AI stack.

The 5x versus 15x productivity comparison alone is worth sitting with. Local, free, and offline gives you meaningful acceleration. Cloud gives you more. The question is whether the gap is worth the cost and the data exposure for your specific situation.

That is a decision worth making deliberately, not by default.

If you are working through questions like this for your own operation, xovionlabs.com is a good place to start.