DeepSeek-v4-flash-vision-exp

DS being unable to precisely view Playwright screenshots is the only thing I really miss from Sonnet. This is promising.

> Images are converted into tokens based on their dimensions, and these tokens are billed together with your text tokens.

> Before inference, every image is automatically resized:

> - Images with a total pixel count below roughly 384×384 are scaled up while preserving their aspect ratio.

> - Larger images are scaled down while preserving their aspect ratio so that the total pixel count after resizing is roughly that of an 800×800 image.

> As a result, there is an upper bound of 384 tokens per image: for example, a 2000×2000 image and a 5000×5000 image consume the same number of tokens after resizing. When a request contains multiple images, each image is counted independently under the same rule—there is no separate calculation for multi-image requests.

400 tokens per image results in 2,500 images per dollar, if I’m not mistaken.

edit: format.

It fails the simple clock test for me which Qwen3.8 27B got (nearly) right. given an image of a clock https://files.catbox.moe/kgwa5e.png

I asked it "what time does the clock show?" (both on reasoning: high)

DS answered: The clock shows *5:10* (and 45 seconds). Here is the breakdown: * *Hour hand (red, shortest):* Pointing at the *5*. * *Minute hand (green, longest):* Pointing at the *2*, which represents 10 minutes. * *Second hand (blue, medium):* Pointing at the *9*, which represents 45 seconds.

Qwen answered: The clock shows *8:10* (with the red second hand on the 5, i.e. *8:10:25*).

- *Hour hand* (short, blue) → 8 - *Minute hand* (long, green) → 2 (10 minutes) - *Second hand* (thin, red) → 5 (25 seconds)

Correct answer is 08:09:25.

I've heard that DeepSeek v4 Flash 0731 has frequently assumed that it has vision capabilities and then resorts to inventing text-based image analysis tools when it finds that it actually can't see. In that case, this is a great upgrade for the model.

Anecdotally, I had to tell 0731 to refrain from viewing screenshots since it kept breaking its sessions by trying to read images.

News announcement with benchmarks: https://api-docs.deepseek.com/news/news260821/

Looks like multimodal training is really useful, app developers might need to consider adapter multimodal agents

> Larger images are scaled down while preserving their aspect ratio, so that the total pixel count after resizing is roughly that of an 800×800 image.

It's useful but for OCR and a lot of other applications it needs to be a bit higher (eg: putting in a full A4 / Letter sized page)

Congratulations! DeepSeek has finally gained eyes — the dark days are about to be behind us.

The benchmark results look promising when compared to Opus 4.8, but for agentic usecases it's lacking images as tool call result types. Giving the model a tool to take screenshots and verify its work is my main usecase for vision models, but this is more oriented towards "build a website that looks like this" type prompts. Hopefully we'll see this by release.

I just ran my image recognition benchmark on it ("is this XXX public landmark"?) and it misses a lot that bytedance seed 2.1 turbo gets right; for example: Asked "Is this Salisbury Cathedral" and supplied a picture of Wells Cathedral, it answers "Yes, the west facade of Salisbury Cathedral". Bytedance seed 2.1 turbo correctly says no. Similar results for a picture of Manhattan Bridge sent as Brooklyn Bridge, Chartres Cathedral sent as Notre Dame, etc. I have a benchmark of 12 such images and seed gets 11/12 and deepseek only gets 6/12.

The DeepSWE benchmark they report (59.3%) overlaps with the confidence interval of 5.6-Sol Medium (61% +/- 2%), but likely at 1/18th the cost (they did not report the DeepSWE benchmark cost, but v4-flash had this cost ratio against Sol Medium).

Interestingly, v4-flash performed several points worse on DeepSWE at 53% +/- 4%. Assuming this result is verified by DeepSWE officially, it would mark a significant advance in Pareto cost/performance on software engineering tasks.

Interesting. Wasn't Deepseek's founder saying that they had explicitly decided not to focus on multimodal models at all and were going text-only because they believed it was enough to achieve AGI?

Is there a way to test it online so that one doesn't have to resort to getting an API key and python code ?

800x800 is 640,000 pixels, or 0.64 Megapixels. That is less than the resolution of computer screens from 1995, Super VGA which has around 0.79 MPs.

This is useful for a reasonable amount of use-cases, but I think the watershed rez will be around triple that, ~1080p, which is enough for almost anything, except small text and subtle details.

Hello Ox Alpha?

For what do you guys use vision in those models? surveillance is the obvious use case... but are there some "nicer" ways to use it?

Was this the ox alpha model?

I main V4 Pro at work now, and at home I route between Pro and Flash based on task. Switched to Opus 4.6 for some tasks at work because I needed image input - horrible. So nice to get image input with DS.

Edit: I see it has limited resolution. Luckily I just built a vision worker plugin for DSH that routes image input to Kimi K2.6 on Cloudflare.

Deepseek flash v4 july sounds like fun and games while you're looking at prices, but it routinely outputs incoherent rubbish and fails to call tools correctly.

Sadly oversold. I hold little hope for the vision model either now.

Benchmarks got a little bump from this: https://xcancel.com/deepseek_ai/status/2087864585504305397?s...

will this be open weights?

[dead]

[flagged]

[flagged]

[dead]

[dead]

[dead]