S
pending
What DeepSeek-OCR's 10x Token Compression Means for Feeding Documents to LLMs
Grounded / Real
Inflated / Uruttu
Original Content
Text is not the cheapest way to get text into a model.
DeepSeek-OCR showed a rendered page costs roughly 10x fewer tokens, and 97% of the words still come back out.
Here's what that changes about how you feed documents to an LLM:
(The numbers are silly when you see them side by side).
1,000 words as text: about 1,300 tokens.
The same page rendered to an image: about 100 vision tokens.
Under 10x compression DeepSeek-OCR decodes 97% of the text correctly.
At 20x it still holds around 60%. You get to pick how much accuracy you trade for room in the window.
This is not just a paper thing. It's already running in stuff you use.
Gemini treats every PDF page as an image and bills it at a flat 258 tokens, no matter how many words are on the page. Glyph (ACL 2026) renders long text to images for 3-4x compression and about 4x faster prefill, which lets a 128K model take on million token workloads.
Computer use agents have been reading screenshots instead of scraped text for a while now.
The idea people are chasing next: downsample old conversation turns to blurrier images, so agent memory gets vaguer with age while it stays in the window.
Where it breaks: anything that needs token level precision. API keys, exact quotes, code you plan to run, long IDs. Keep those as text and compress the prose around them.
Try it out. Take a doc you keep pasting in, screenshot the pages, send those instead, and compare the token count against the answer quality.
Would you let your agent keep a blurry version of turn 3, or does fuzzy memory break the whole thing for you?
DeepSeek-OCR showed a rendered page costs roughly 10x fewer tokens, and 97% of the words still come back out.
Here's what that changes about how you feed documents to an LLM:
(The numbers are silly when you see them side by side).
1,000 words as text: about 1,300 tokens.
The same page rendered to an image: about 100 vision tokens.
Under 10x compression DeepSeek-OCR decodes 97% of the text correctly.
At 20x it still holds around 60%. You get to pick how much accuracy you trade for room in the window.
This is not just a paper thing. It's already running in stuff you use.
Gemini treats every PDF page as an image and bills it at a flat 258 tokens, no matter how many words are on the page. Glyph (ACL 2026) renders long text to images for 3-4x compression and about 4x faster prefill, which lets a 128K model take on million token workloads.
Computer use agents have been reading screenshots instead of scraped text for a while now.
The idea people are chasing next: downsample old conversation turns to blurrier images, so agent memory gets vaguer with age while it stays in the window.
Where it breaks: anything that needs token level precision. API keys, exact quotes, code you plan to run, long IDs. Keep those as text and compress the prose around them.
Try it out. Take a doc you keep pasting in, screenshot the pages, send those instead, and compare the token count against the answer quality.
Would you let your agent keep a blurry version of turn 3, or does fuzzy memory break the whole thing for you?
Validated Content
Confirmed
DeepSeek-OCR: 10x compression → 97% OCR precision, 20x → ~60% — matches the paper exactly (arXiv 2510.18234).
1,000 words ≈ 1,300 tokens as text vs. ~100 vision tokens rendered — consistent with the paper's own framing.
Gemini bills every PDF page as a flat 258 tokens regardless of word count — confirmed directly from Google's Gemini API documentation.
Glyph (ACL 2026): 3-4x compression, ~4x faster prefill, lets a 128K model handle ~million-token workloads — confirmed against the published paper and ACL Anthology listing.
Computer-use agents reading screenshots instead of scraped text — accurate, standard characterization of how these agents work.
Softer/speculative (not false, just not a hard citation)
"Downsampling old conversation turns to blurrier images" as the next idea people are chasing — reasonable extrapolation (the DeepSeek-OCR paper itself gestures at "memory forgetting mechanisms" as a research direction), but it's framed as an emerging trend rather than something documented/measured, so it's more speculative than the rest.
Bottom line: This post holds up very well — nearly every hard number and technical claim traces back cleanly to primary sources (the DeepSeek-OCR paper, Gemini's official docs, and the Glyph ACL paper). The only soft spot is the closing "what people are chasing next" line, which is plausible but not something I could verify as an active trend, just a logical extension of the research.
DeepSeek-OCR: 10x compression → 97% OCR precision, 20x → ~60% — matches the paper exactly (arXiv 2510.18234).
1,000 words ≈ 1,300 tokens as text vs. ~100 vision tokens rendered — consistent with the paper's own framing.
Gemini bills every PDF page as a flat 258 tokens regardless of word count — confirmed directly from Google's Gemini API documentation.
Glyph (ACL 2026): 3-4x compression, ~4x faster prefill, lets a 128K model handle ~million-token workloads — confirmed against the published paper and ACL Anthology listing.
Computer-use agents reading screenshots instead of scraped text — accurate, standard characterization of how these agents work.
Softer/speculative (not false, just not a hard citation)
"Downsampling old conversation turns to blurrier images" as the next idea people are chasing — reasonable extrapolation (the DeepSeek-OCR paper itself gestures at "memory forgetting mechanisms" as a research direction), but it's framed as an emerging trend rather than something documented/measured, so it's more speculative than the rest.
Bottom line: This post holds up very well — nearly every hard number and technical claim traces back cleanly to primary sources (the DeepSeek-OCR paper, Gemini's official docs, and the Glyph ACL paper). The only soft spot is the closing "what people are chasing next" line, which is plausible but not something I could verify as an active trend, just a logical extension of the research.