DS is dead to me after the pricing changes. Qwen 3.8 has replaced it entirely for me
DeepSeek-v4-flash-vision-exp
161–169 of 169 posts
Re: DeepSeek-v4-flash-vision-exp
#162Earlier quoted context omitted.
Well it succeeded on first try for me: https://i.imgur.com/gljOYr9.png [Image 1] what time does the clock show? + Thought: 368ms The clock shows 8:10. - Blue hour hand: just past 8 - Green minute hand: pointing at 2 (10 minutes) - Red second hand: pointing at 5 (25 seconds) ▣ Plan · DeepSeek V4 Flash Vision Exp · 4.4s edit: I now ran the test 10 times total, fresh session every time, and DS4 got it right 9 out of 10…
Try it on openrouter directly if you don't believe me that i got this result. Maybe image inputs behave differently in opencode, or it could even be that this agent harness uses different models for different tasks, not sure. Or maybe the reasoning (high) makes the difference.
I don't doubt you. I just think it's good practice to run more than n=1 before conclusions are made. We're all engineers here.
Re: DeepSeek-v4-flash-vision-exp
#163DS is dead to me after the pricing changes. Qwen 3.8 has replaced it entirely for me
isn't Qwen more expensive? Much more verbose than flash on high, and and slower?
Re: DeepSeek-v4-flash-vision-exp
#164I think my only gripe with DeepSeek was the lack of vision capabilities. This is an awesome and much awaited update!
Re: DeepSeek-v4-flash-vision-exp
#165I 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…
Do you publish your benchmark somewhere? What's the best vision (captioning) AI you've come across?
Re: DeepSeek-v4-flash-vision-exp
#166Earlier quoted context omitted.
You say "it can be trained" but they fail at counting in my use cases, let alone maintaining symbolic relationships. Can you point me at one that can understand a detailed block diagram? Frontier is fine, soliciting recommendations
For counting, there are specialized counting models, e.g. https://huggingface.co/spaces/MengqiLei/count-anything-demo I tried to parse hand-drawn ER diagrams in the past and did not have much success with any model, frontier or otherwise. If you have annotated data, I'd recommend finetuning a recent (dense) VLM, but don't expect 100% accuracy. https://unsloth.ai/docs/basics/vision-fine-tuning
good advice; tbh I'm really trying to use counting as a proxy for "understand symbol, flag annotation next to it, associate annotation with symbol and store as object in location".
Ideally, look at second diagram and see something similar in the same location and understand it's the same physical object.
Re: DeepSeek-v4-flash-vision-exp
#167I 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…
Do you publish your benchmark somewhere? What's the best vision (captioning) AI you've come across?
My benchmark is tiny compared to WorldVQA or FG-BMK, which are available, so I'd point you in that direction if you're interested in a VLM benchmark. My use-case isn't exactly captioning as in "what is in this image?" -> caption, I am using the VLM to validate captions, as in "is this an image of [supposed subject]?" - my ranking of models I've benchmarked is gemini-3.7-flash > seed-2.1-turbo > gpt-5.6-luna > qwen-3.7-plus > qwen-3.7-flash. Gemini is almost perfect on my test dataset, only failing on some esoteric pop-culture minor celebrities and being over-specific in some cases (i.e. Q: is this [common name of fruit]? A: that's a [latin species name of fruit], not a [common name of fruit]; false). However, gemini-3.7-flash is only in my test list because openrouter has it on 75% introductory discount; otherwise it would be about 4x more expensive than seed.
Re: DeepSeek-v4-flash-vision-exp
#168flash weights shipping a vision model that outruns last year's flagship is the most reliable beat in this field. the frontier isn't a ladder anymore, it's whoever recompiled last night.
Re: DeepSeek-v4-flash-vision-exp
#169The 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…
The closer comparison would be 5.6-Luna. On DeepSWE at Xhigh it's 57% at 1/6 the cost of Sol M, on Max 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.
Luna is a very capable model - thanks for pointing that out. Terra is the strange one: not cheap enough or intelligent enough to be on the frontier. But Luna sure is.