DeepSeek's Files API: Upload Images Once, Reuse Them by ID
DeepSeek's Files API lets you upload an image once and reference it by ID across requests, avoiding repeated uploads and wasted tokens.
What is DeepSeek’s Files API?
DeepSeek’s Files API is a way to upload an image to DeepSeek’s servers once and get back a short file ID, which you can then reference in as many API calls as you want instead of re-sending the raw image data every time. It shipped alongside DeepSeek V4 Flash Vision (Experimental), the company’s newly released vision-capable model, and it solves a specific waste problem: sending the same base64-encoded image or file bytes over and over for every follow-up question you ask about it.
TL;DR
- DeepSeek released V4 Flash Vision (Experimental), its first vision-enabled model in the V4 Flash line, which can read photos, screenshots, charts, and documents alongside text.
- The Files API lets you upload an image a single time and reuse the returned file ID across multiple requests, rather than resending the full image data with every call.
- Images are capped at 384 tokens each and billed at the same rate as V4 Flash text, meaning vision requests currently cost about the same as text-only ones.
- You can supply images three ways: as base64, as a URL, or through the Files API, and all three methods work across chat, message, and response-style API calls.
- Uploaded files can persist for up to 30 days or permanently, and the upload itself is free, so the only cost you’re managing is the token cost of the model actually reading the image.
- Independent testing showed strong performance on charts, financial tables, and stylistic art analysis, but real mistakes on dense math notation and multilingual handwriting, so the model isn’t flawless despite the hype around new releases.
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
Why does re-uploading images waste tokens and bandwidth?
Without a files mechanism, every time you want to ask a model a new question about an image, you have to send that image’s full data again, whether that’s a base64-encoded blob or a hosted URL the model has to fetch. If you’re asking five different questions about the same chart or document, you’re paying the upload and processing overhead five separate times. For large images or documents, that’s repeated bandwidth, repeated latency, and in some API designs, repeated token costs for the image portion of the request.
This becomes especially wasteful in workflows where you’re iterating: asking a model to first extract text from an image, then asking it to summarize that text, then asking a follow-up question about a specific detail. Each of those calls, under the old approach, would carry the entire image payload again even though nothing about the image changed.
How does DeepSeek’s Files API work?
The workflow is straightforward. You upload an image to DeepSeek’s file storage endpoint one time. DeepSeek stores the file and returns a file ID, a short string that acts as a pointer to that stored image. From that point forward, instead of attaching the raw image to your request, you just reference the file ID.
This means you can fire off multiple distinct prompts, asking different questions entirely, while pointing at the same uploaded image. In a demonstrated test, a single photo was uploaded once, generating one file ID, and then two separate questions (one asking where the photo was likely taken, another asking about the ecosystem shown in the image) both referenced that same ID without re-uploading the file. Both responses came back correctly, and the image itself was only transmitted once.
The Files API sits alongside the two other supported input methods, base64-encoded images and direct URLs, and all three are supported across DeepSeek’s different API interaction styles (chat completions, messages, and responses). Uploading a file is free, and files can be retained for up to 30 days or set to persist permanently, giving developers flexibility depending on whether they’re testing something short-lived or building a reference library of images an application will query repeatedly.
What can DeepSeek V4 Flash Vision actually do?
V4 Flash Vision is DeepSeek’s first attempt at giving its V4 Flash model image understanding. It accepts photos, screenshots, charts, and documents directly alongside text prompts, and each image is capped at 384 tokens, billed at the same rate as ordinary V4 Flash text.
In hands-on testing shown in a recent review, the model handled some tasks impressively and stumbled on others:
- Dense math notation: Asked to transcribe a complicated equation into LaTeX, the model got the overall structure right but made a symbol-placement error, misreading an exponent as part of a different number.
- Multilingual handwriting: Given a handwritten sample mixing English, Urdu, Arabic, and Indonesian, the model handled English and mostly handled Arabic and Indonesian, but dropped the Urdu text entirely and appeared to fabricate a word in the Indonesian portion that wasn’t actually present in the image.
- Business chart analysis: The model performed strongly here, correctly reading three overlapping data series, pulling exact figures from the chart, and reasoning about causality, for example identifying that one metric peaked a full year before another and correctly characterizing a shift from a shortage regime to a correction regime.
- Dense financial tables: Given a low-quality four-column financial table, the model reconstructed the structure, subtotal markers, footnotes, and negative values in parentheses with only minor numeric errors.
- Art and style analysis: Given an old master-style drawing, the model correctly identified the artistic medium, the historical style, and the specific artist, grounding its answer in visual details actually present in the image rather than generic guesses.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
The pattern that emerges is a model that’s genuinely capable at visual reasoning and structured data extraction, but still prone to hallucination on tasks requiring precise character-level transcription, especially across less common languages or symbol-dense notation.
Is the Files API worth using?
For any workflow that asks multiple questions about the same image, yes. If you’re building something like a document Q&A tool, a chart analysis assistant, or any application where a user uploads an image once and then interacts with it conversationally, the Files API removes the need to resend that image on every turn. Given that the upload itself is free and files can persist for up to 30 days or indefinitely, there’s little downside to using file IDs over raw base64 payloads in any multi-turn scenario.
For one-off, single-question use cases, the benefit is smaller since you’re only making one call anyway. In those cases, base64 or a direct URL is simpler to implement and skips the extra upload step entirely.
Frequently Asked Questions
What is DeepSeek V4 Flash Vision?
It’s DeepSeek’s newly released, experimental vision-capable version of its V4 Flash model. It can process images (photos, screenshots, charts, documents) alongside text prompts, with each image capped at 384 tokens and billed at the same rate as text.
Does the DeepSeek Files API cost extra money?
No. Uploading a file through the Files API is free. You still pay the standard token cost for the model processing the image content itself, but the act of storing and referencing the file by ID carries no additional charge.
How long do uploaded files stay available?
Files uploaded through the API can be retained for up to 30 days, or set to persist permanently, depending on how the upload is configured.
Can I still send images as base64 or a URL instead of using the Files API?
Yes. DeepSeek supports three input methods for images: base64-encoded data, direct URLs, and the Files API. All three work across DeepSeek’s chat, message, and response API formats.
Is DeepSeek V4 Flash Vision reliable for tasks like OCR or transcription?
It’s strong on structured data like charts and tables, and capable of genuine visual reasoning rather than simple pattern matching. However, testing has shown it can make mistakes on dense mathematical notation and multilingual handwritten text, including dropping entire languages or inserting words not present in the source image. Since it’s an experimental release, results should be verified rather than trusted outright.
