alt.hn

8/21/2026 at 10:33:56 AM

DeepSeek-v4-flash-vision-exp

 https://api-docs.deepseek.com/guides/vision/

by dares2573

8/21/2026 at 2:04:24 PM

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.

by leumon

8/21/2026 at 2:40:54 PM

I’ll keep that in mind next time I need to tell what time it is by asking an llm to read an analog clock.

Snark aside, I’m not sure that these gotcha tests are any more useful than asking politicians gotcha questions. Sure, the model can’t tell me what time it is, but it can code the Wang algorithm for noisy audio matching in one shot. Maybe this is just me being an optimist, but this is my hiring philosophy and I guess maybe now my llm philosophy: I’m not interested in seeing how dumb I can make you look, I’m more interested in how smart you can be.

by dghlsakjg

8/21/2026 at 2:30:11 PM

This is not a normal looking clock - most clocks have either one color for all hands (second hand is thinnest and maybe also longest) or one color for hour/minute and one for second. I know that the hand lengths and thicknesses on this image are correct but for some reason I, a totally human person who grew up when analog clocks were still common, see this and think the hand on the 5 is the minute hand. How does the AI do if you just make all the hands black?

by emosenkis

8/21/2026 at 2:39:16 PM

then deepseek answers: "The clock shows 8:25. The short hour hand is pointing to the 8, and the long minute hand is pointing to the 5, which represents 25 minutes."

and qwen still answers: "The clock shows *8:10* (with the second hand on the 5, i.e., 25 seconds). - *Hour hand* points to the 8 - *Minute hand* points to the 2 (= 10 minutes) - *Second hand* points to the 5 (= 25 seconds) So the time is *8:10:25*, or simply *8:10*."

by leumon

8/21/2026 at 2:43:09 PM

Qwen still got the wrong answer, though.

Are we more forgiving because it’s the same type of mistake a human would make?

by dghlsakjg

8/21/2026 at 2:20:08 PM

A good share of humanity would have also gotten this question wrong!

by ttul

8/21/2026 at 2:31:05 PM

It's been four years that we are looping those

"The professional failed its task!" // "Laymen would have failed it too".

Which makes no sense.

by mdp2021

8/21/2026 at 2:22:51 PM

Yeah, I heard most kids these days can't read analog clocks either.

I can't actually remember where I learned to read a clock, it might have actually been in school. I guess that means they don't teach it anymore. (Everyone's phone shows the time anyway...)

by andai

8/21/2026 at 2:11:11 PM

Gemini 3.7 Flash and 5.6-Sol (on all reasoning levels) also answer 8:10:25. The new "stealth" Ox Alpha also replies with the same. Opus 5 replies with 8:10 (no seconds). Not sure why this is so hard for them; Gemini is especially good at vision and I would have expected better from it.

by ComputerGuru

8/21/2026 at 2:11:15 PM

welp, damning indictment. not sure if that means DS is super crap, or qwen is super good

by nubg

8/21/2026 at 2:22:12 PM

Neither. Performance of all models is incredibly spikey.

by wolttam

8/21/2026 at 11:08:44 AM

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.

by ciberado

8/21/2026 at 11:11:21 AM

Oof 800 by 800 kills a lot of use cases

by knollimar

8/21/2026 at 11:33:30 AM

Might still be fine. The most recent crop of vLLMs proactively use whichever programs are available on the system (e.g. ImageMagick or PIL) to "zoom in" by cropping subimages if they can't quite make out the details.

by johndough

8/21/2026 at 2:03:07 PM

Downsizing a higher res image to lower res means the zoom will be blurry.

by knollimar

8/21/2026 at 2:19:35 PM

They process the original image file with Python on the local device. (And I've seen the web chats do this with their "computer use" features too.)

The really wild one is even blind models will do this and they'll try to run stats on the pixels to figure out what it looks like... the even wilder thing is that it kind of works!

by andai

8/21/2026 at 11:31:54 AM

For most use cases you can fix that in the harness. Just give the model a tool to request a crop of specific coordinates of any image it has in its context. Call the tool "zoom" and it should be intuitive for the model

Maybe there are some use cases where you need high detail everywhere at once, but for OCR of small text and the like a zoom ability should be sufficient

by wongarsu

8/21/2026 at 11:57:12 AM

For really dumb models I've also had success automatically cropping it into a grid of N images with the max size, then processing each cell individually, then once all been processed, do one final call with resized image + all other context previously generated per cell. Basically a workaround to the image dimension restrictions without loosing fidelity. Works well with even dumb 7B models.

Can't remember if I stole this idea from some existing public harness though, can't remember. If someone knows of public harnesses that do this already, please share them :)

by embedding-shape

8/21/2026 at 12:40:10 PM

Does this not loose context? Especially e.g. in fonts where the character pairs 0O 1I 1l Il may be difficult to differentiate?

by dotancohen

8/21/2026 at 12:57:12 PM

That's what the grid crop should handle. The detail is retained at that level, and then everything is logically stitched together again using the lower-res-full-image as reference. That's going to be 2x token usage at minimum though.

by skeledrew

8/21/2026 at 11:40:22 AM

It might also be due to its experimental status. Wouldn't surprise me if the GA version allows for larger input. Either that or the eventual pro version.

by shadyr

8/21/2026 at 1:50:01 PM

what are these use cases?

by Chnmy

8/21/2026 at 2:04:16 PM

Anything where there are symbols representing in space (e.g. schematics). Thats pretty broad

by knollimar

8/21/2026 at 11:18:04 AM

flash vs fine details. Pick one.

by asdfsa32

8/21/2026 at 11:32:24 AM

Gemini "flash" models have an option for media resolution, including a high resolution option for screenshots.

by Doohickey-d

8/21/2026 at 12:53:00 PM

At what price point?

by skeledrew

8/21/2026 at 1:08:43 PM

[flagged]

by 13639366668

8/21/2026 at 2:28:30 PM

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.

by ttul

8/21/2026 at 2:33:09 PM

The closer comparison would be 5.6-Luna. On DeepSWE at Xhigh it's 57% at 1/6 the cost of Sol M, on Mac it's 67% at 1/3rd the cost.

Still an advance, I just thought it worthy to note Sol isn't nearly as impressive on the cost/performance frontier as discounted Luna.

by paytonjjones

8/21/2026 at 11:00:54 AM

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.

by LorenDB

8/21/2026 at 11:45:01 AM

It tried to recreate vision by analyzing pixels on 3 separate projects I had it working on.

by VulgarExigency

8/21/2026 at 1:52:15 PM

I've mitigated this by giving it a "skill" that just means the harness using a different model.

by trollbridge

8/21/2026 at 2:00:33 PM

Yeah I've seen it a lot. It goes through the effort, unasked, of pulling screenshots off a connected device and then it's like... Oh shit yeah I can't see.

by mavamaarten

8/21/2026 at 2:20:50 PM

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.

by jerkstate

8/21/2026 at 11:31:40 AM

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

by BrucecarlL

8/21/2026 at 12:59:14 PM

Or about to start. Depending on which life philosophy you desire to believe.

by doublerabbit

8/21/2026 at 2:31:52 PM

Im intrigued. Please do share these philosophies.

by unified101

8/21/2026 at 11:06:58 AM

> 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)

by zmmmmm

8/21/2026 at 11:12:24 AM

Can split and feed?

by mkagenius

8/21/2026 at 11:17:20 AM

that's difficult as well, how do you k ow where to split?

by throwaw12

8/21/2026 at 12:23:24 PM

Text is often written as separate lines (and paragraphs) at least in some languages.

by kgwgk

8/21/2026 at 11:47:34 AM

Let the model do the splitting. A 800x800px image should be enough to make those decisions

by wongarsu

8/21/2026 at 11:43:48 AM

Overlap the splits?

by grog454

8/21/2026 at 11:40:09 AM

Presumably a small cheap model could do that part?

by vrganj

8/21/2026 at 1:37:42 PM

Was this the ox alpha model?

by Johnny_Bonk

8/21/2026 at 2:38:37 PM

That would've been a very strange arrangement.

by WiSaGaN

8/21/2026 at 11:29:47 AM

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.

by gozucito

8/21/2026 at 11:32:09 AM

You'd expect a tool-enabled model to leverage crop and zoom tools to inspect and validate what it thinks it's seeing, though.

by barrkel

8/21/2026 at 11:47:32 AM

I typically provide small screenshots to llms so this seems fine for that usecase, providing an entire screens context seems cause confusion with a lot of llms.

by dakolli

8/21/2026 at 1:11:46 PM

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

by wiz21c

8/21/2026 at 1:31:21 PM

You can use the playground on openrouter. Still needs an account and some money, but it's one of the more useful accounts to have sitting around with a $5 of balance. Great for one-off experiments with various models

https://openrouter.ai/deepseek/deepseek-v4-flash-vision-exp

by wongarsu

8/21/2026 at 12:07:42 PM

Hello Ox Alpha?

by erikkri

8/21/2026 at 2:14:26 PM

Nope. Handles vision differently.

by ComputerGuru

8/21/2026 at 12:23:07 PM

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?

by 5kyn3t

8/21/2026 at 2:40:24 PM

https://stencil.so/blog/snapcompact - some agents (notably oh my pi, i forget which others) come with snapcompact as a primary means of compaction. Take the entire context, stick it in a small font in a PNG, and vision capable models can summarize and pull out the most useful information in many fewer vision tokens than the original context used.

I've not used it myself, but it's there.

by zdragnar

8/21/2026 at 12:25:54 PM

The obvious use case, especially on HN, is frontend dev of any kind at all. The second most obvious one is OCR of paper documents.

by deaux

8/21/2026 at 12:32:34 PM

Frontend Dev? I do not really understand. do you let the models analyze the webpages you are working on? or for testing?

by 5kyn3t

8/21/2026 at 1:35:34 PM

LLMs are not great at aligning stuff on first try, they are however very good at taking screenshots and fixing their mistakes. Claude Design also does this all the time, as does regular Claude in the web UI if you tell it to make a powerpoint presentation

I really missed this feature when I had DeepSeek code a small game for fun. When writing UI and rendering code it could execute the game and get screenshots back, but then had to rely on my feedback on what had gone wrong. Models with vision can do much better here, finding more issues on their own

by wongarsu

8/21/2026 at 12:44:57 PM

Standard flow with a vision model in OMP is to write the front end code, fire up the server, fire up a headless browser and then take screenshots and examine and iterate. Works great. When I'm using DeepSeek V4 Flash, it always reminds me instead that I have to validate manually by loading up the page.

by rpdillon

8/21/2026 at 12:38:45 PM

QA of course. You hook up your agent with CDP access to live product + let it screenshot and look into result. Also you could hook agent with CDP access to Figma to read/write, there a vision model is very useful as well.

by dandaka

8/21/2026 at 12:56:46 PM

It closes the development loop. Without it a model can't check if the stuff it made actually visually renders like it's supposed to. It can only guess/assume.

by deaux

8/21/2026 at 12:35:15 PM

but for OCR there are much better suited models, I use mlx-community/PaddleOCR-VL-8bit

by dandaka

8/21/2026 at 12:55:40 PM

Sometimes you intentionally want to verbatim keep "mistakes", sometimes you don't and want them to be "fixed". OCR-only models tend to only do one of those two, in VLM cases often the latter. With multi-modal LLMs you can just tell them (adherence of course needing evals/differs per model).

by deaux

8/21/2026 at 1:55:44 PM

Having vision is very handy for getting it to make plots/figures with matplotlib. A model with vision can be much more autonomous with catching visual glitches/misalignments and correcting itself.

Also used it for 3d printer control once, had it diagnosing issues, calibrating my Tradrack MMU and canceling failed prints autonomously from a couple of cameras placed around the printer.

by hgoel

8/21/2026 at 1:53:35 PM

Allowing it to analyse a system under test (usually in an emulator, web browser, Electronic app container, etc. - something that can be reasonable captured).

It makes running much, much longer feedback loops possible. Although you can mix and match non-vision and vision models simply by invoking a vision model when you need one, as I like to use non-vision models like glm-5.3.

by trollbridge

8/21/2026 at 1:51:14 PM

In the feedback loop when working on anything UI or graphical output related.

by kzrdude

8/21/2026 at 1:11:58 PM

Any kind of spatial/graphical task is likely going to go better with a vision-capable model. Feed it a napkin-sketch of what your app should look like. Have it verify screenshots of the UI it just built. All of these one-shot-a-video-game evaluations that have suddenly become popular only work if the model can interpret screenshots...

by swiftcoder

8/21/2026 at 12:37:13 PM

My product is connecting employers and workers with conversational agents. They love to communicate with images — CVs, documents, photos of worksites. Even CV-as-photo or offer-as-photo format is very popular. My daily driver Deepseek Flash can't see those photos. So I use image models to let agents understand the context.

by dandaka

8/21/2026 at 12:40:48 PM

I've been working on an agentic graphic design tool, so vision is quite useful for having the model check its own work. I'm already seeing improvements with this model vs the text-only one.

by moonu

8/21/2026 at 1:53:33 PM

No one’s mentioned robots, so… robots. VLA models, etc.

by wolttam

8/21/2026 at 1:42:42 PM

Generating alt text for images in social media posts.

by dcre

8/21/2026 at 1:26:19 PM

Going straight to surveillance and unable to think "nicer" ways... is strange.

1. process graphs and charts

2. process handwritten math formula, also chinese characters writings

3. process design sketch and wireframe

4. process scanned documents

... etc

in fact these transformer models currently suck for surveillance, too slow and expensive. There are already faster and better facial/gait/object recognition models out there.

by dudisubekti

8/21/2026 at 2:22:16 PM

when i am learning i draw what i undestand in a picture and ask ai to correct me. i want ai to watch over me while i am learning.

this is such good way to learn something for me.

by dominotw

8/21/2026 at 12:31:24 PM

[dead]

by MagicMoonlight

8/21/2026 at 11:23:06 AM

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?

by v9v

8/21/2026 at 12:10:01 PM

It was explicitly said that they are pursuing multimodal support. A quote from the meeting transcript: https://github.com/demo-zexuan/liang-wenfeng-investor-meetin...

    Nevertheless, as a component, we will undoubtedly implement multimodal support — and we are already doing so. We plan to develop relevant models, ensuring that versions like V4 and subsequent iterations will natively support multimodal functionality.
Earlier, the following was said, which might match more what you had in mind.

    Achieving excellence in AI training does not require a global model or even multimodal approaches—by narrowing the scope of AI training and eliminating multimodality, certain tasks may remain unachievable without compromising the algorithm's validity.

    Multimodal approaches ultimately need to be implemented.
It is difficult to tell who said what, since the speaker ids are missing.

by johndough

8/21/2026 at 1:13:24 PM

Worth noting that deepseek has had a separate vision-capable model for some time, which also powers their chat interface's vision mode

by swiftcoder

8/21/2026 at 11:48:52 AM

I think you're thinking of Dario saying this about image generation.

by dakolli

8/21/2026 at 11:59:46 AM

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.

by try-working

8/21/2026 at 11:17:49 AM

will this be open weights?

by dsrtslnd23

8/21/2026 at 11:57:19 AM

I believe so. Openness has always been a consistent tradition of DeepSeek

by dares2573

8/21/2026 at 12:22:40 PM

I imagine this is based on their 'Thinking with Visual Primitives' paper, and they had mentioned that the weights would be released for that

by moonu

8/21/2026 at 11:41:17 AM

This is something I would like to know as well.

But if not, does anybody know a recommended way to attach vision to deepseek flash (on a self-hosted infrastructure)?

by griffiths

8/21/2026 at 2:35:19 PM

[flagged]

by promptsphere

8/21/2026 at 12:17:26 PM

[flagged]

by locitra

8/21/2026 at 11:15:52 AM

[dead]

by lzy

8/21/2026 at 12:30:17 PM

[dead]

by MagicMoonlight

8/21/2026 at 11:40:42 AM

[dead]

by jaksdbvqi37u