DeepSeek V4-Flash Vision: What It Nails and Where It Fails
Hands-on tests of DeepSeek's first vision model show strong chart reading and OCR but real mistakes on handwriting and math symbols.
What is DeepSeek V4-Flash Vision?
DeepSeek V4-Flash Vision (labeled experimental, or “EXP”) is DeepSeek’s first vision-language model built on top of V4-Flash. It accepts images alongside text through the same API, capping each image at 384 tokens and billing at V4-Flash’s text rate. It can take images as base64 data, as a URL, or through a new Files API that lets you upload an image once and reuse it by ID across multiple requests, avoiding the cost of re-sending the same file every time. Independent testing shows the model is strong at chart interpretation and structured document extraction, but inconsistent on handwriting and prone to small errors in dense math notation.
TL;DR
- Vision was bolted onto V4-Flash rather than shipped as a separate architecture, and DeepSeek prices image input at the same rate as text, which is unusually cheap for a multimodal model.
- Chart reading and reasoning is the standout skill: in testing, the model pulled exact figures off a three-series business chart and correctly identified a leading indicator relationship a full year ahead of the trend it predicted.
- Dense financial tables come out nearly perfect, with subtotal markers, footnote superscripts, and negative-value formatting preserved even from a low-quality scanned image.
- LaTeX transcription of a complex equation was structurally correct but included a subtle exponent placement error, a reminder that visual math OCR still needs a human check.
- Multilingual handwriting recognition was the weakest area, with the model dropping an entire language from a mixed-script sample and inserting words that weren’t in the original text.
- The new Files API removes redundant uploads by returning a reusable file ID valid for up to 30 days, cutting token costs for repeated questions about the same image.
- Creative and visual-reasoning tasks performed well, including a full narrative writing test and correctly identifying the artist and style of an old master drawing from visual details alone.
Seven tools to build an app. Or just Remy.
Editor, preview, AI agents, deploy — all in one tab. Nothing to install.
How does DeepSeek V4-Flash Vision actually work?
The model extends the existing V4-Flash text model with a vision encoder, so images and text flow through the same request format across chat, completions, and response-style API calls. Each image is tokenized at a fixed cap of 384 tokens regardless of resolution or complexity, which keeps costs predictable but also means very dense images are compressed into a limited representation before the model reasons over them.
Pricing follows the same rate as V4-Flash text tokens, which DeepSeek had recently raised. That makes image input relatively inexpensive compared to vision models that charge a premium per image, though the overall session cost in testing (image-heavy, multi-turn) still came out to roughly $2 for a full run of test prompts.
Underlying model files published for the vision variant show it built on a MoonViT-style vision tower paired with the V4-Flash MoE backbone, distributed across dozens of safetensors shards and packaged with quantized (NVFP4/FP8) formats aimed at newer accelerator hardware, alongside an SGLang deployment path for self-hosting.
How good is it at reading math and equations?
Tested against a dense, messy handwritten-style equation with a prompt asking for exact LaTeX transcription, the model got the overall structure right, correctly parsing nested terms and formatting. But it made at least one meaningful error: it rendered an exponent as part of a multi-digit number instead of keeping it as a separate superscript, turning what should have been “S2 to the power 4” into something read as “S24.” That’s a small mistake visually but a significant one mathematically, since it changes the meaning of the expression. The takeaway for anyone using vision models for equation OCR: treat the output as a strong first draft, not a final answer, especially with exponents, subscripts, and symbol placement.
Can it read handwriting and mixed languages?
This was the weakest result in testing. Given a handwritten sample mixing English, Urdu, Arabic, and Indonesian, the model handled English cleanly, including punctuation and spacing, and mostly handled the Arabic and Indonesian text. But it dropped the Urdu portion entirely, producing no transcription for that script at all, and it inserted a word into the Indonesian transcription that didn’t appear in the source image. That’s a hallucination, not just a misread. For any workflow depending on multilingual handwriting extraction, particularly scripts beyond Latin characters, this suggests the model needs a verification pass rather than blind trust, even though the video creator noted this is expected for a first-generation experimental release.
How well does it interpret charts and business data?
Other agents ship a demo. Remy ships an app.
Real backend. Real database. Real auth. Real plumbing. Remy has it all.
This is where the model performed best. Given a business chart with three overlapping data series spanning a supply-chain disruption period, it correctly read all three series, extracted specific figures directly from the chart (a 26-week peak, an 8x drop, a $620 billion figure, and an 18% decline), and went further into actual reasoning. It identified that one metric (lead time) peaked roughly a year before another (exports) peaked, correctly framing it as a leading indicator relationship rather than just describing two separate trends. It also correctly characterized a shift from a shortage-driven period to a correction period later in the timeline. That’s closer to genuine chart reasoning than simple value extraction, and it’s the kind of capability that matters for anyone using vision models to summarize dashboards or reports.
Is it reliable for document and table extraction?
For a dense, low-quality four-column financial table, the model reconstructed the structure with high fidelity: section groupings, subtotal markers, footnote superscripts, dashes, and parenthetical negative values were all preserved correctly. There were minor numeric errors here and there, but the overall accuracy was estimated at roughly 97 to 98 percent against the source image, which is strong performance for a scanned, imperfect document. For OCR-style extraction of structured tables, this is a genuine strength.
What about creative and visual reasoning tasks?
Beyond OCR-style tasks, the model was tested on a long-form creative writing prompt involving a character forced into a high-stakes emotional decision, with structural constraints that had to be followed exactly. It produced a coherent, in-character narrative that respected the prompt’s constraints and included a self-assessment score when asked. It was also given an image of an old master pen drawing and asked to identify the artist, medium, and style. It correctly described the visual details (a crowned figure holding a cross, sword, and orb, wearing fur-trimmed robes), identified the medium as pen and the style as German Renaissance, and named Albrecht Dürer as the likely artist, referencing a monogram detail on the sheet as supporting evidence. That’s a case of the model grounding a conclusion in specific visual evidence rather than pattern-matching a generic description.
Frequently Asked Questions
What is DeepSeek V4-Flash Vision EXP?
It’s DeepSeek’s first vision-capable model, an experimental extension of V4-Flash that accepts images alongside text input through the same API, priced at the same rate as text tokens.
How much does it cost to use?
Image input is billed at the same per-token rate as V4-Flash text, with each image capped at 384 tokens. A multi-prompt testing session covering OCR, charts, tables, and creative writing cost roughly $2.
What is the Files API DeepSeek released alongside it?
It’s a free feature that lets you upload an image once and receive a reusable file ID, valid for up to 30 days, instead of re-sending the full image data with every request. This cuts token usage for workflows involving multiple questions about the same image.
Is DeepSeek V4-Flash Vision good at OCR?
It’s strong on structured content like tables and charts, reconstructing complex layouts with high accuracy. It’s weaker on handwriting, particularly multilingual samples, where it can drop entire languages or insert text that wasn’t in the original.
Should I trust it for math transcription?
Treat it as a strong starting point rather than a final answer. In testing it got the overall structure of a complex equation right but misplaced an exponent, changing the mathematical meaning of that term.
