Local AI by the gigabyte

Shopping lists by RAM tier, all-in-one products, and a Redis creator's 14-hour days

Salvatore Sanfilippo (@antirez on X), the Italian programmer who created Redis (the database that almost every large website uses to store temporary data), has spent the last week working 14 hours a day. He has been building DwarfStar 4, free software for running a top open-weights AI model on a high-end Mac. People are calling it DS4. In a short blog post, he says DS4 has unexpectedly taken off. For the first time since starting with local AI, he is reaching for the model on his own machine for serious work. He uses it instead of Anthropic’s Claude or OpenAI’s ChatGPT.

Translated for a general reader: a well-respected programmer who has used hosted AI services for years has switched. Some of his daily work now happens on his laptop, with no internet connection involved. He still thinks the hosted services are stronger, but the gap has closed enough that the trade is worth it.

This article was reviewed and hand-edited, using curated sources I hand-picked, drafted by AI using personalised style guidance. I find it useful to keep in touch with the dozens of X.com bookmarks I come across every few days. Maybe you do too.

A shopping list by RAM tier

Until recently, advice on running AI models on your own machine was scattered across tweets, GitHub issues, and Discord servers. This week a developer who goes by jun_song posted a tidy shopping list by memory tier. It functions as the consumer-electronics buying guide local AI never had.

The headline tiers, in plain English:

• Around 32 GB of memory: a small Google-flavoured open-weights model.

• 32 to 64 GB: a mid-sized open model from Alibaba, in the Qwen family.

• 96 to 128 GB: 2 Chinese ones, DS4 and MiniMax.

• 256 GB: a sizeable model from Xiaomi.

• 512 GB: a very large model that needs most of half a terabyte sitting in memory.

A 96 GB Mac currently runs to about £4,000. A 128 GB one is closer to £5,500. The 256 and 512 GB tiers do not exist as a single laptop, only as desktop boxes (or stacks of graphics cards) that cost as much as a small car. The practically relevant tier for most people remains the bottom three.

What is new is that the list is short and confident. A year ago, recommending a single model for a given memory bracket would have been irresponsible. Everything moved week to week. The pace has slowed enough that “buy this model for this much RAM” is now a sensible thing to say out loud.

In the replies, one reader asked the question generalists actually care about: which tier reaches the quality of Claude or Google’s Gemini? Jun’s answer was that the 32 to 64 GB bracket is not it. The Mac that gets you to “comparable to the hosted services” is still in the £4,000 and up zone.

From models to products

The more interesting move is what Sanfilippo did with DS4. DS4 is integrated software wrapped around one specific model, with one goal: be the easiest way to use a particular open-weights AI on a high-end Mac.

This matters because it inverts the usual local-AI experience. Until now, running a local model meant a stack of choices. You picked the model, the runtime (the software that loads and executes it), and a chat interface. You configured file paths and memory limits. You accepted that any of those layers might break next week. DS4 collapses the stack into one thing that comes with sensible defaults.

Sanfilippo is upfront that the project will move on from the current model when something better comes along. He also floats a more interesting idea: separate local models tuned for specific work. A coding version, a legal version, a medical version. You load whichever one fits the task at hand.

That is a different shape from how hosted AI works today. ChatGPT and Claude try to be good at everything. A local stack with a 128 GB ceiling cannot afford that. Specialisation is forced on you by the hardware. The bet I would take is that forced specialisation will end up producing better tools for specific jobs than a generalist that has to cover every base.

His sign-off line is unusually direct for a developer blog: “AI is too critical to be just a provided service.”

The agent-friendly plumbing

Underneath the products, the plumbing has been getting tidier too. A developer who goes by bstnxbt shipped a new release of a Mac inference runtime called dflash-mlx this week. The headline note: it is “much more usable for real OpenCode / coding-agent sessions”. OpenCode is an open-source coding assistant, similar in style to the ones offered by Anthropic and OpenAI.

Translated: a coding assistant pointed at this runtime can hold a longer conversation about your codebase without slowing to a crawl. It also keeps “what we already worked out” notes around, so it does not start from scratch each turn. It also reuses earlier output when you ask for similar code again, which saves time.

This is not the kind of change that makes a YouTube video. It is the unglamorous side of local AI becoming a working tool. Caches that survive across sessions, retries that do not corrupt state, tool calls that actually stream back without dropping characters. A hundred small tool calls in a row need to land without the runtime falling over.

🔮 Prediction

The next twelve months of local AI will be defined by products. Expect three or four DS4-style integrated bundles to emerge, each pairing one specific open-weights model with a polished runtime and a clear use case. Coding will go first, because the audience is technical, motivated, and willing to spend £4,000 on a laptop. Writing assistants and document analysis will follow. The shopping guide will get shorter. Instead of “pick from the 47 models that run on your RAM”, it will read “if you want coding, install this; if you want long-document work, install that”.

What would change my mind? Two things. If the next hosted-model release widens the quality gap noticeably, the integrated-local-product story stalls. The same applies if Apple’s memory pricing stops falling. In either case we go back to “hosted services for everything that matters.” Neither feels likely on this week’s evidence.