Can an AI PC write your resume offline?

What did Microsoft just launch, and does it matter for your CV?

Microsoft launched the Surface Laptop Ultra on 7 October 2026: a Windows laptop built around Nvidia's RTX Spark N1X platform that can run AI models of 120 billion parameters or more locally, starting at $2,599 and shipping 16 October. Satya Nadella and Jensen Huang presented it together in San Francisco, which tells you how much both companies are betting on moving AI work off the server and onto your desk. The chip pairs an Nvidia Grace CPU — 18 or 20 cores depending on configuration — with a Blackwell RTX GPU carrying up to 6,144 cores and up to 128GB of unified memory. Microsoft quotes roughly one petaflop of peak AI compute on select configurations, though that figure is theoretical FP4 performance with sparsity rather than a real-world number.

The software side is where it gets interesting for anyone writing a job application. Microsoft also showed Hybrid Intelligence, a GitHub Copilot feature that routes a prompt to an on-device model when the local model can handle it and falls back to the cloud when it can't, plus MAI Code 1.1 Flash — an on-device model with 137 billion total parameters and 6.8 billion active, running at 3-bit quantization with a 256K context window. Microsoft recommends 120GB or more of memory to run it. Those are developer tools, not CV tools. But the plumbing underneath — a laptop that holds a capable model in memory and answers without a round trip — is exactly what a private CV workflow needs.

Here's the honest bridge to your job search: none of this was built for job seekers, and you don't need a $2,599 machine to get most of the benefit. What the launch does is normalise something that used to be a hobbyist setup. On-device AI is being sold as the grown-up way to handle sensitive documents, and a CV is the single most sensitive document most people will ever hand to a chatbot — your full name, your street address, your phone number, every employer you've had and, increasingly, your salary expectations. If you've hesitated over that upload button, what your CV file actually exposes when you paste it into an assistant is worth reading before you decide.

Microsoft also claims the laptop returns local AI prompts about twice as fast as a MacBook Pro with an M5 chip, with image generation 4.3x quicker and video 6.2x quicker than an M5 Pro. Treat those numbers carefully — the company hasn't published a test methodology for any of them. And the broader AI PC story is still early: analysts don't expect these machines to dominate laptop shipments until around 2028. Nvidia-based devices from Asus, Dell, HP and Lenovo arrive on the same 16 October date, so prices will move. For a CV, raw speed matters far less than whether the model in front of you writes well.

Can a local AI model actually write a decent CV?

A local model will write a competent CV, not a brilliant one. Models in the 8B to 14B range handle the mechanical work well — tightening a rambling bullet, cutting a 900-word job description down to four lines, converting duties into achievements, matching the vocabulary of a posting you've pasted in. That's most of what CV editing actually is. Where local models still fall behind the big cloud systems is judgment: deciding that your 2019 career break should be framed as consulting rather than hidden, or that the project you're proudest of is the one a hiring manager will care least about. Bounded rewriting is solved. Strategy isn't.

You don't need Microsoft's flagship to try this. Two free desktop tools cover almost everyone: Ollama if you're happy in a terminal, and LM Studio if you'd rather click. LM Studio has no cloud component at all, which makes it the simpler choice when privacy is the whole point, and it lets you run two models side by side on the same CV draft — genuinely useful for seeing which one writes like a human. Sizing is straightforward: roughly 8GB of RAM for a 3B model, 16GB for a 7–8B model, 24GB of VRAM once you reach the 30B class. A mid-range laptop from the last three years will run something usable.

The trap is tone. Small models reach for safe, generic phrasing, which is how you end up with a CV full of "spearheaded cross-functional initiatives" and "results-driven professional" — the exact register recruiters have learned to skim past. A local model also has a shorter memory of your voice across a long session, so by the third pass your bullets start sounding like each other. Read every line aloud before you export. If it doesn't sound like something you'd say in an interview, cut it, because you'll have to defend it there. If you're unsure whether your draft has drifted into machine phrasing, it's worth checking whether your CV reads as AI-written before you send it.

Does on-device AI really keep your salary and address private?

Yes, and this is the strongest argument for the whole approach. When a model runs on your own hardware, your prompt never travels anywhere — there's no API call, no server-side log, no retention window to read about in a terms-of-service page. Microsoft's own on-device language model for Windows, Phi Silica, is documented as processing prompts and generating responses entirely on the device, with nothing transmitted to Microsoft or a third party during inference. Run a model through LM Studio or Ollama in local-only mode and the same holds. You can paste your real address, your real notice period and your real current salary without stopping to redact anything first.

That changes the practical workflow more than it sounds. Most careful job seekers strip identifying details before pasting a CV into a chatbot, which means the assistant is reasoning about a document that isn't quite yours — and then they have to stitch the real details back in by hand. Offline, that step disappears. The same logic drives the wave of private assistant launches, and what on-device processing genuinely changes when you hand an assistant a full CV is the same question whether the model sits on a phone or a $2,599 laptop. The answer is consistent: local inference removes the upload risk, not every risk.

Two caveats deserve your attention. First, "AI PC" doesn't mean "offline Copilot". Hybrid features are explicitly designed to hand a prompt to the cloud when the local model can't cope, and several Copilot experiences still need a Microsoft account and a connection. If offline processing is the reason you bought the machine, confirm which model is answering before you paste anything sensitive. Second, your laptop becomes the single point of failure. A model that keeps your salary history on local disk has also put your salary history on a device that can be stolen, shared or backed up to a cloud folder by default. Encrypt the drive.

What can't an offline model tell you about passing an ATS?

A local model cannot see your file. It sees the text you typed into a chat window — not the PDF you're about to upload, not the order a parser extracts that PDF in, and not what happens to your two-column layout when the text boxes are read left to right across the page. That gap is the single biggest limitation of writing your CV offline, and no amount of local compute fixes it. A model can tell you a bullet is weak. It can't tell you that your job titles ended up interleaved with your dates, or that your skills section was extracted as one unbroken string of characters.

It also has no recruiter-side data, and that matters more than people assume. An offline model has never seen which version of your CV got a callback, how your experience reads against the other 300 applicants for the same posting, or what a particular employer's screening setup actually indexes. It's working from general writing knowledge, so it will confidently approve a CV that's well written and still badly positioned. Ask a local model whether your CV is ATS-friendly and you'll get plausible, generic advice about keywords and simple formatting — correct in principle, untethered from your actual document.

Then there's the accuracy problem nobody mentions in the launch keynotes. Local models hallucinate the same way cloud ones do, and on a CV a hallucination isn't a harmless slip — it's a fabricated metric you have to defend in an interview, or a job title that doesn't match what your former employer will confirm. Quantisation makes this slightly worse: squeezing a model to 3 bits to fit in memory costs something, even when benchmark scores hold up. Check every number, every date and every title the model produces against your own records. Treat the output as a draft written by a fast, well-read intern who has never met you.

How should you combine offline drafting with a real CV check?

Split the job in two: draft offline, verify on the file. Use the local model for everything that involves your private details and your voice — rewriting bullets, compressing a decade into one page, pasting the full job ad and asking which of your projects maps to which requirement. Then export the actual PDF or Word document you intend to send, and have that file checked as a file. This is the step people skip, and it's where applications quietly die. A CV can be beautifully written and still arrive at the recruiter as scrambled text, because the thing being screened is the extraction, not your prose.

That verification step is what HRLens is built for. Upload the PDF, Word file or even a photo of a printed CV — Hebrew OCR included — and the free analysis returns an overall score out of 100, five category scores covering experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility, plus your strengths and a visual layout analysis that shows how the page is actually structured. Paste the job description too and it lists the skills you're missing against that specific posting. You can run your finished CV through a free analysis and see where the file itself is working against you.

My honest opinion on the AI PC story: the privacy gain is real, and it's also solving the second-biggest problem. The thing standing between most applicants and an interview isn't a data leak — it's a CV that parses badly, repeats the same three verbs, and says nothing measurable. An on-device model gives you a private, tireless editor, and that's worth having. It does not give you the recruiter's view of your document. Do both: write where nobody's watching, then check the file against the system that will actually read it. One of those is free, and neither takes an afternoon.

Frequently asked questions