Can Muse Spark Open review your CV offline on a Mac?

What did Meta and Apple actually launch on 15 October 2026?

On 15 October 2026, Meta released Muse Spark Open, an open-weight model that Apple's updated MLX framework can run locally on Apple silicon. The specs are the headline: 816 billion total parameters in a mixture-of-experts design, with roughly 42 billion active on any given step, a one-million-token context window, and native multimodality — text, images, audio and video go in, text comes out. Meta shipped two checkpoints, a full BF16 version and a lower-precision NVFP4 variant that trims the memory footprint. Mark Zuckerberg publicly credited Apple's John Ternus for the engineering work behind the Mac support. The first Muse Spark landed back in July 2026; this is the open-weight sibling.

For job seekers, the interesting word in that announcement isn't "parameters" — it's "open-weight". Open weights mean the model file is downloadable, so inference can happen on hardware you own, with nothing leaving your machine. That matters because the most common objection to AI CV tools is the one you've probably had yourself: your CV carries your home address, your phone number, your salary trajectory and a decade of employer names, and handing all of it to a server you can't inspect feels careless. A model running on your own silicon is the first clean answer to that worry — an architectural answer rather than a reassuring paragraph in a privacy policy.

Two caveats before you get excited. Open weights don't mean free in practice — you pay in hardware, electricity and setup time instead of per token. And "runs on a Mac" is doing an enormous amount of work in that sentence, as the next section explains. Muse Spark Open is also served through a long list of hosted API providers and cloud model catalogues, which is how the overwhelming majority of people will actually touch it. The second you go through one of those, your CV is back on someone else's infrastructure under whatever retention terms they publish. The privacy win belongs to local inference specifically, not to the licence.

Meta's benchmark claims put it ahead of competing open models on agentic and visual reasoning tasks, and the demos leaned hard into coding: fixing bugs in a Mastodon client, driving a video editor to review footage. Note what's missing from that list. Nothing about documents, hiring, or recruitment screening. The model is strong at long multi-step reasoning, which is genuinely useful when you feed it a 900-word job ad and ask it to map your experience against every requirement. But raw capability on coding leaderboards tells you nothing about whether a recruitment system can read the PDF you're about to send.

Can Muse Spark Open actually run on the Mac you own?

Almost certainly not. Running Muse Spark Open locally takes roughly a terabyte of unified memory, which on Apple hardware means two to four Mac Studios wired together. The configuration Meta and Apple showcased pairs two M5 Ultra Mac Studios with 512GB each, connected over Thunderbolt 5, with MLX pooling the memory so the two machines behave as one. Apple hasn't published pricing or a firm ship date for the 512GB Mac Studio beyond late October. The alternative — four 256GB M5 Ultra Mac Studios — comes in around $38,000, given that a single 256GB machine starts at $10,799. That is a server rack with a nicer aluminium finish.

Here's what your actual Mac can do. With 4-bit quantisation, an 8B-class model occupies about 4–5GB and runs comfortably on a 16GB MacBook Air at somewhere between 12 and 40 tokens per second depending on the chip. A 14B model wants around 9GB and a 24GB machine. A 30B-class model needs roughly 18–20GB of weights, so 32GB of unified memory is the realistic floor — and you should keep weights under about 60% of total RAM so there's headroom for the KV cache and for macOS itself. Sparse mixture-of-experts models are kinder: a 35B model with 3B active can load in about 22GB.

So the headline is true and irrelevant to you at the same time. That's the honest verdict, and it's worth stating plainly because a lot of coverage glossed over it. The genuinely good news is that you don't need 816 billion parameters to tighten a bullet point or catch a tense inconsistency — that work sits well within reach of a model that fits on a mid-range laptop. If you want the longer version of this argument applied to Windows hardware, the same ceiling shows up in what on-device AI really changes for CV drafting, where the bottleneck turns out not to be model size at all.

What can an offline model genuinely fix on your CV?

Language work, and it's better at it than most people expect. A local model running on your own Mac will rewrite a flabby bullet into something specific, fix tense and number agreement across twelve years of roles, flatten inconsistent date formats, cut the word count on a summary that's drifted to six lines, and brainstorm the vocabulary a particular job ad is reaching for. Feed it the ad and your current bullets side by side, and ask which required skills you've demonstrated but never named. That exercise alone surfaces gaps on most CVs — you did the work, you simply described it in your old employer's internal vocabulary instead of the industry's.

It's also the right tool for the drafts you'd be embarrassed to upload anywhere. Early-career CVs where you're still deciding whether a student society counts. The version with your real salary expectations typed at the bottom. The one where you're writing honestly about why you left. A local model gives you a sounding board with no retention window, no training-data clause and no account. That's not a theoretical concern: audits of resume tools have repeatedly found content retained for months after account deletion, with some policies allowing anonymised CV data to improve the vendor's own models.

Where small local models fall down is predictable. They pad. Ask for a punchier bullet and an 8B model hands back something longer with three adjectives bolted on. They invent metrics — a bullet about "improving onboarding" comes back as "reduced onboarding time by 40%", and if you don't catch it you've just put a fabrication in front of a hiring manager who may check. They also reach for exactly the buzzwords recruiters have learned to skim past. The same failure mode shows up on the privacy-first assistants now shipping, which is covered in what on-device processing really protects.

What can't a local model check, no matter how big it is?

It can't tell you how your file behaves when a parser opens it. This is the blind spot that size never fixes — an 816-billion-parameter model reading pasted text is working from the clean version, not from the PDF you'll actually attach. Two-column layouts get read in the wrong order or interleaved into nonsense. Text boxes and SVG icons are often skipped entirely, so the "Skills" panel you're proudest of simply doesn't exist downstream. Contact details parked in a header or footer frequently vanish. Tables collapse. None of that is visible to a model that only ever saw the words.

Nor can a local model calibrate you against a market. It has no idea how many applications a candidate with your profile needs to land an interview in your city, what your experience pays, or which of your five CV sections is dragging the rest down relative to other applicants for that role. It will happily give you a confident score out of ten, and that number is generated from nothing. This is the same ceiling that bigger hosted models hit too, as the comparison of Meta's two Muse models on CV work found: more reasoning, same structural blind spot.

Layout analysis needs the document rendered, not described. Whether your margins squeeze the page, whether your font falls apart at the size you've chosen, whether page two orphans a single bullet — those are visual properties of the artefact. A text-only conversation can't see them, and a multimodal model looking at a screenshot sees what a human sees, not what an extraction pipeline produces. The practical fix is to stop asking one tool to do both jobs and run the file itself through a free ATS resume check before you send it anywhere.

How do you keep your CV private and still get a real ATS check?

Split the work. Do your thinking, rewording and honest first drafts on the local model, then de-identify the file and run that version through a tool that actually inspects the document. You lose almost nothing by doing this, because the information an ATS compatibility check needs is structural — headings, date formats, column count, embedded objects, font encoding, section order — and none of it depends on your street address being present. Replace your real contact block with a placeholder, keep everything else intact, and the parsing verdict you get back applies perfectly to the file you'll send.

Be specific about what comes out. Strip your full street address down to a city, remove your date of birth, any national ID or tax number, your personal mobile if you're jumpy about it, and the names of referees who haven't agreed to be listed. Leave in employer names, job titles, dates, technologies, qualifications and metrics — those are the things any check needs to reason about, and they're already on your public professional profile anyway. If a tool asks for more than that before giving you a result, that's your signal to look elsewhere.

That's the gap HRLens is built to close. Upload the de-identified PDF, Word file or even a photo of your CV — Hebrew OCR included — and the free CV analysis returns a score out of 100, five category scores covering experience, tech stack, impact and ownership, clarity and structure, and ATS compatibility, plus a visual layout analysis of the actual document. Paste the job description and you'll see which required skills are missing, and every analysis generates ATS-friendly rewrites across six templates that mirror the ad's keywords. Your local model can then polish the wording offline. That division of labour beats either tool alone.

Frequently asked questions