Loading today’s stories
Product1 distinct publisher3 min readPublished
Wired's walkthrough puts free Meta and Google models on a laptop with 8 GB of RAM at the low end, which makes the no-upload option real for teams with sensitive files, as long as somebody owns the updates.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
Follow any of these and your For You feed starts watching them — no settings page required.
build
Inco AI's DFlash 2: 21% longer accepted drafts for 1.3% latency and 18.5M parameters1 distinct publisher
build
Meta's real announcement is the split: 30B on your GPU, everything else behind the API6 distinct publishers
product
A billion downloads, and nobody will say what a download is1 distinct publisher
build
Intel puts its Arc GPU operating knowledge inside the coding agent already installed1 distinct publisher
The decision actually gets made on a screen called Get local models. In LM Studio Bionic you create a project, give it a name, open the prompt box, and land in a picker that lists each model with its size, its popularity, a short description, and staff picks for people who do not want to choose [12]. The first tradeoff a new user meets there is disk space rather than quality: the smaller files download faster and take up less room, and they do less [13]. Behind that picker sits Hugging Face, which lists more than three million models [11]. Nobody in a six-person practice is evaluating three million of anything, which makes that staff-picks row the most consequential piece of UI in the flow.
Wired frames capability as already settled rather than as the open question. Its own framing is that the free downloads are not as advanced or as speedy as what sits inside a paid app, and are capable enough for everyday use [4]. Everyday use means what a paralegal or a clinic administrator needs done to a document they are not allowed to upload, rather than the demo you give a partner.
Hardware is where the honest cost shows up. Between the 8 GB floor and the 32 GB the guide says the biggest and fastest models require, there is a factor of four in memory [14], and a gap that size is a purchase order rather than a preference. Wired also points at VRAM above 8 GB on a discrete GPU, and a dedicated Nvidia card if you are on Windows [8]; macOS gets the nod because Apple Silicon puts the CPU, GPU and memory together [6]. There is no published minimum spec [9], which in practice means the machines you already own decide the model you can standardize on.
The guide frames this as a maintenance exercise rather than a model-selection one: updates get handled by you [5]. And because it is a setup walkthrough, what a real evaluation would need is not in it: a retention curve, usage depth, whether the local app is still being opened in week six [15]. A pilot worth running measures time to a useful answer on one named document workflow, then checks a month later whether the same person is still doing it.
The forcing function has two axes. Can the file leave your network, and does a named human own the updater? File cannot leave, updates owned: local is the honest answer, and the monthly payment and the usage ceiling both go away [2]. File cannot leave, nobody owns updates: you have a pilot that will quietly go unpatched, and it should not reach a fleet. File can leave, updates owned: you are paying salary to reproduce a hosted service, which only pencils out when the metered ceiling is the thing that hurts [2]. File can leave, nobody owns updates: keep the subscription.
The person this is for is specific. Somebody who has already said no to a cloud tool for one particular file, and who can name the colleague who will click update in March.
Ranked by verification strength, evidence, and original report placement.
Wired reports that the large language models powering bots like ChatGPT and Gemini can be run locally on your own computer, with the key benefits being offline access and greater privacy because you are not sending anything to the cloud for anyone else to analyze or review.
Running a model locally means not paying an AI company a monthly subscription and not hitting usage rates.
Numerous LLMs are available to download for free, including from big names like Meta and Google.
There is more maintenance involved in a local setup and you lose some of the convenience of just loading up the ChatGPT app; for example, you need to handle updates yourself.
Local LLMs run on Windows, macOS and Linux, with macOS the preferred platform for most AI enthusiasts partly because Apple Silicon chips combine the CPU, GPU and RAM together, which AI models like.
On memory, the bare minimum is 8 GB of RAM, which limits the size and speed of models you can run; 16 GB is better; and 32 GB or more is required to use the biggest and fastest models.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 29, 2026
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One desk, checkable steps, folklore numbers
Everything traces to a single Wired walkthrough, and it splits cleanly in two. The procedural half — the runner names, the Create Project and Get local models path, the picker showing size and staff picks — is the kind of detail a reader disproves or confirms in ten minutes, which is real evidentiary discipline. The half that decides whether the project works, the 8/16/32 GB tiers and the 8 GB VRAM line, arrives with no model, no quantization and no speed measurement behind it, and Wired itself says there is no minimum spec.
Nobody counted the installs
There is no uptake figure in this reporting to score. The single number that looks like scale — Hugging Face's three million models — measures shelf space, not people running anything, and the piece never touches how many readers finish a setup like this or still use it a month later.
Upside stated, friction whispered
The gap is small and one-directional. Privacy, zero subscription and offline use are asserted as settled goods; the recurring cost — someone has to own the updates — is a parenthesis in the fourth paragraph. What keeps the number low is that Wired volunteers the capability shortfall against paid apps instead of pretending free weights match them.
Evergreen how-to pull, no vendor voice
No company speaks in this piece and nothing in it traces to a press release; the products it names are free, so there is no obvious sale being made. The pressure that does shape it is format pressure — service journalism rewards a confident recommendation, and 'generally considered the best' does that work without naming who considers it or what was tested.
Solid on the how, thin on the whether
Confidence sits mid-range for an unusual reason: the subject is uncontroversial and the instructions are cheap to test, so the risk of being wrong about the steps is low. The risk of being wrong about fit is not, because one publisher's rule of thumb is the only thing standing between a reader and a slow, disappointing model on an 8 GB laptop.