Compress a PDF Without Uploading It, on Your Own Mac

Each service publishes something about retention. Here it is as written on 9 October 2026: what each one states, not a verdict on any of them.

Three details are easy to miss. Smallpdf’s one-hour figure applies to account users; for anyone uploading without logging in, its notice promises only a reasonable period after the file was last opened. Adobe states no clock at all: its page ties deletion to whether you sign in. And CloudConvert’s two pages give different timing, immediately on one and a 24-hour outer limit on the other.

TinyPNG is not in the table because it takes images, not PDFs. It matters if you also shrink phone photos of IDs online: its FAQ says it retains files for a maximum of 48 hours, and its terms say deleted within 48 hours.

A deletion window is a promise about what happens after the upload. What you can observe for yourself is narrower: whether the bytes left your Mac. If an engagement letter, a vendor policy or a client contract limits which third parties may hold the file, that is a question about the transfer, and with a local app there is no transfer to ask about. The product-by-product detail is in our Smallpdf, iLovePDF, CloudConvert and TinyPNG comparisons.

Every portal sets its own, and a firm or lender may configure a different limit from the vendor’s default. These are named examples read at source. For your own, search the portal’s help page for “file size” or ask the lender’s processor, then aim at 70% of the number. Our upload-limits guide gives the reason: portals count the encoded upload and some re-wrap the file, so a file just under the cap can fail. That is advice, not something we measured.

Among the tight caps are forms that feed another institution: the IRS and lender platforms at 15 to 20 MB, besides the Greenhouse parser (2.5 MB) and Exchange email (10 MB) above. Portals built for accounting firms are generous, in gigabytes, though Canopy’s page shows the catch: size can switch features off long before it blocks the upload. The IRS figure comes from a news release, and the IRS notes that such items may not be updated after release. Its current upload-tool page lists JPG, PNG and PDF but states no size, so confirm the figure on the live upload screen.

Email is the quiet constraint. Encoding adds about 37% to an attachment, so a 25 MB cap means aiming at 18 MB (18 × 1.37 = 24.7 MB), as the email guide works out. Courts keep their own list, which differs from these; it is in our page for law firms and in the legal and HR guide.

One rule is not about size. The EEOC portal says to upload no password-protected files, so the file’s owner has to remove the password before you compress it. We have not tested how Smol handles a password-protected PDF, so have the owner remove the password first and treat any other result as unverified.

Smol’s PDF presets are long-edge pixel caps, not DPI targets: 600, 900, 1,200, 1,800 and 2,400 px for Tiny, Small, Medium, Large and Original, with JPEG quality of roughly 25, 45, 60, 75 and 88. The same preset therefore lands at different resolutions on differently sized pages; Medium measured 200 PPI on one document and 109 on another. Original is not lossless. It re-encodes.

Two cautions. A preset name tells you nothing about legibility, so open the output at 100% on the page with the smallest print. And for phone photos use Compress rather than Convert: converting ignores the quality setting and writes a large JPEG. Converting HEIC to JPEG also carried GPS coordinates through in all three routes tested in the HEIC conversion guide, Smol’s convert included. Raise the maximum dimension, which goes up to 32,768 px, when small print is the point of the document. The sibling guides on scanned PDFs and what a photo leaks beyond its metadata cover each in detail.

Make the text layer first, then shrink the pictures. A scan is a photograph of a page, so the only searchable text is whatever OCR adds, and OCR needs enough pixels to read. In our test, an OCR engine read 100% of the words at 109 PPI (Medium), 96.8% at 82 PPI (Small) and 1.42% at 55 PPI (Tiny) when the scan was compressed before any OCR. Run OCR on the original, as our scanned-PDF test does, and every preset kept all 1,405 words searchable.

A statement is mostly figures, and that test measured words in a services agreement, not digits. Even with OCR done first and the text layer intact, Small leaves the page image at 82 PPI, and a digit that is hard for a person to read is a risk in a balance. The same scan, compressed to Small before any OCR, was read at 96.8% of the words, so even on prose some characters were lost, and we did not measure digits. The page you check at 100% before sending is the one with the totals, and Small is a last resort for a tight cap. The same guide recommends 150 PPI or above for body copy and says to scan at 300 PPI, so you have room in both directions.

If the bank issued the statement as a PDF download, it is not a scan at all. It is text, the presets have no images to work on, and the file is probably under every cap above already. That case, and the scanned transcripts and forms that school portals reject, is covered on our education page.

1. 1

   Put the outbound set in its own folder

   Smol writes a copy next to each original with a _compressed suffix and does not overwrite by default. A separate folder per matter or period still keeps what goes out apart from what stays.

2. 2

   Find the cap and take 70% of it

   Use the portal’s help page, or ask the processor. Against the 15 MB in the IRS fact sheet that is a target near 10 MB per file.

3. 3

   OCR the scans before compressing

   Smol does not OCR. Use your scanner software, OCRmyPDF or Acrobat, and keep the order: compressing an un-OCRed scan to a low resolution leaves nothing to recognize. In our test, one OCR text layer on a synthetic scan survived every Smol preset; we did not test layers made by scanner software or Acrobat.

4. 4

   Drop the files into Smol and pick a PDF preset

   Start at Medium for scans. Mixed batches are fine: PDFs and photos go in together.

5. 5

   Read the per-file results

   Check each output size against your target. On a direct compress, if the result would be larger than the file, Smol keeps the original. Open the smallest-print page of one output at 100%.

6. 6

   Search one output before you send it

   Open an output, press Cmd-F and search a word you can see on the page. If it finds nothing, the text layer did not survive, so do not send that file.

7. 7

   Run Strip metadata on the outbound copies

   Do this after compressing, because compression does not remove properties. On its own the strip is reversible: one command from a free command-line tool restored the cleared properties in our test.

8. 8

   Rewrite each outbound file with qpdf

   Run qpdf --linearize outbound.pdf outbound-clean.pdf on every copy. We measured that rewrite after exiftool’s strip, not after Smol’s, and we did not test the full order of compress, strip and rewrite in one run.

9. 9

   Check comment authors, attachments and filenames by hand

   The strip does not touch them. The metadata section below covers why.

Do not take it from a web page, ours included. Three checks, cheapest first.

1. 1

   Watch Activity Monitor

   Open Activity Monitor, choose the Network tab and find Smol’s row. Run a batch of non-sensitive PDFs and watch the sent-bytes column. macOS is doing the reporting, not the app. Our published screenshot shows that row at zero bytes sent while files are processed.

2. 2

   Turn the network off

   Disconnect Wi-Fi and Ethernet and run the same batch. If it finishes, the compression does not depend on a server.

3. 3

   Use a per-connection firewall

   Smol’s privacy notice names Little Snitch and LuLu for this. It lists three kinds of request: a license check on first launch and then weekly, which sends the license key and a hashed identifier for the Mac; an update check, which sends the app and macOS versions; and crash reports, which are off unless you opt in. It says the only domains reached are smolmac.com and its subdomains. So expect to see those. Traffic that appears when files are processed is what you are looking for.

The same three checks work on a browser-based compressor that says it runs client-side, with one difference. Load the page, then disconnect and drop in a non-sensitive file: if it compresses, the work happens in your browser. Or open the browser’s developer tools on the Network tab and look for a request that carries the file. Neither check settles trust, because the page’s script can change on the next load, while an installed app stays at one version until you update it.

That is a statement about the app on your Mac, at the version you tested (1.0.35 when this page was written). It is not a certification, an audit or a statement about any law or contract. We make no claim of compliance with any regulation or standard, and whether this satisfies your firm’s vendor policy or your clients’ agreements is a question for your compliance lead. We cannot answer it.

Compression does not remove it, and the strip operation removes less than the word suggests. On a synthetic three-page contract (162,058 bytes, invented names) the Tiny preset gave 57,377 bytes, 64.6% smaller, with all 10 Document Info keys and 13 XMP tags intact. The full test compares seven routes.

Reviewer names matter to HR and audit teams: a policy circulated for comment carries the name of everyone who commented, and that is document content, so no metadata tool removes it. The published test found that a strip alone can be undone, and that a rewrite is what made it hold. The rewrite is one command, run after the strip:

qpdf --linearize outbound.pdf outbound-clean.pdf

That post ran the rewrite after a strip made with exiftool, and the recovery attempt then found nothing to restore. For Smol’s strip, one command from a free command-line tool restored the cleared properties in the same test, and we have not measured a rewrite after Smol’s strip, so treat Smol’s strip alone as reversible. The post also found that rewriting first and stripping second left a reversible file, and that compression collapsed the file’s update history while stripping left it. That is why we do not call PDF metadata removal permanent. Nor is it redaction: a black box drawn over a name in Preview leaves the text selectable. For folders, see batch-stripping metadata.

For a single document, Preview or the Terminal route is usually enough, and if your team already pays for Acrobat Pro, its redaction and OCR are reasons to keep using it. We say so in why Preview’s filter ruins some PDFs and in compressing from the Terminal. The wider field is in the PDF compressor roundup and the PDF Squeezer comparison.

Team is $49 once for three Macs, with seats managed from an admin dashboard. Enterprise is $20 a seat once, for 5 to 500 seats, from one dashboard. Solo is $29 once for one Mac. Updates are free and there is no per-file quota. There is no free trial and purchases are not refundable. The only free thing is the browser demo on the homepage, which is not the app. Judge it from the published measurements, or buy one Solo license ($29) and test your own scans before buying for an office.

• Write down the preset settings the team agrees on and recreate the preset on each Mac, so every preparer starts from the same settings. A preset is saved on the Mac where you make it.
• Add the Finder Quick Action so a right-click compresses the selection from Finder.
• A watched folder suits fresh image exports, such as phone photos of IDs or receipts. Do not point one at your scanner’s output folder: OCR has to come before compression, and a watched folder compresses unattended. Nor at files that were already compressed, because a known defect can make optimized files larger.

For a whole folder of metadata strips, our timed comparison favors exiftool on speed: 4,030 ms for 90 mixed files against 15,151 ms for Smol, which handled 80 of them and refused the other ten, because the strip did not accept video or audio in that test.

Seven cases, and the first three are the ones that matter most.

You need redaction. Smol does not redact, and clearing metadata is not redaction. Use a dedicated redaction tool and verify by extracting the text from the output, not by looking at it. The Acrobat alternatives guide is a place to start.

You need a documented chain of custody. If every step has to be hashed and written down, that is a records workflow, not a drop target. Smol shows per-file results rather than a log of who touched what.

You are cleaning one document and secrecy is the point. exiftool -all= followed by qpdf --linearize is free, takes a second, and is better than the strip we ship.

Your PDFs are text. Payroll registers, bank exports and statements from accounting software have no images for the presets to work on. Ghostscript did better on our text file, and ReportLab-made PDFs compress by almost nothing.

The portal demands PDF/A or a signature must survive. We make no promise that output stays PDF/A, so check it in a validator. A signature covers the exact bytes of the file, so compress before signing, not after.

The Mac never touches the network. The license is checked weekly and the app keeps running offline for up to 30 days without reaching the license server, according to Smol’s privacy notice. A permanently air-gapped machine needs a plan for that.

You are on Windows, or you want the file processed in a pipeline. Smol is macOS only and runs as an app on one Mac, not as a step in a server pipeline.

Where a local app earns its $29: a month of mixed scans and photos in one pass, originals left alone by default, and nothing sent to anyone. See also the real estate page, where disclosure PDFs raise the same file-size questions, and the law firm page, where court filing limits differ from the ones above.