Hacker Newsnew | past | comments | ask | show | jobs | submit | fzysingularity's commentslogin

The author’s comments on vision providers is especially interesting. We saw that most providers don’t provide native video url support, have high-variability in vision performance (likely due to the fact that they’re serving different quantization levels behind the same model id).

If you’re building vision-native apps, there are so many footguns in vLLM/SGLang serving configurations, let alone the routing/orchestration in providers like OR, that leave the user more confused about the model’s capabilities.


Do you have any advice on this front? I use vllm for a project but only for text tasks at the moment.

text is mostly beaten to death, so you can expect good defaults to work for vllm. VLMs specifically are quite sensitive to quantization, especially if you want it to do fine-grained localization (time or spatial), and the vllm default params can be way off for your use-case.

For example, most vision models don't need 256K context length for modes like Qwen3.8-27B when all you care about is single-image captioning, so you can technically save on KV cache. Video reasoning does require that context length, so it's a different set of deployment parameters that need to be enabled.

All of this to say that the providers that offer these models, are simply using vLLM / SGLang, and mostly cater to the text inference use-case (coding, etc). Vision always seems to be a bit of an afterthought.


Got it, I'll test more deployment params when I benchmark models for vision tasks.

Do you have any tips for Gemma 4 31B in particular? I quite like the model but I feel like I'm underutilizing my rented GPU hard due to skill issues. Throughput should be way higher than this, right?

---

Gemma 4 31B NVFP4, vLLM 0.29, single B200 (modal), FlashInfer, fp8 KV, prefix caching, 32k ctx

Workload: ~6k-token shared prefix (~98% cache hit) + short input, ~250 tokens out, ~250 seqs running. Get ~5-6k output tok/s, flat from 128 to 384 concurrency.

Spec decode (DFlash, n-gram) didn't give a good boost which is annoying because I feel like this task should be easy for a draft model to predict.

Individual request latency doesn't matter I just need as much throughput as possible. Its only active for a few hours when I need it.


Have you tried any vision tasks with this model? We've been serving these on our gateway [1], and the quants are quite terrible for vision. Curious to hear your experience.

[1] https://www.vlm.run/gateway


I'm surprised this post isn't getting as much attention as it should. Crazy times!

VLM Run (https://vlm.run) | 1x Founding Infrastructure Engineer

We’re building the inference platform for visual intelligence. We’re a deeply technical team of veteran AI / computer-vision engineers (20+ years combined, MIT/CMU PhDs) who’ve shipped production ML infrastructure across autonomous driving and LLMs.

We just launched the VLM Run Gateway (https://vlm.run/gateway), a unified API that serves open-weight VLMs, embodied VLAs, ViTs, served across a fleet of GPUs and clouds. That's exactly the infrastructure problem this role will own.

If you’re interested, email us at hiring@vlm.run with your GitHub profile and link to projects you’ve recently built - especially with docker, k8s, GPUs.


I think we all ought to look at the ZDR fine-print here.

I get that in principle that there's no retention, but these are powerful models that can comprehend, paraphrase and summarize your logs for the sake of "product" improvement. Who knows what's collected here.


Frontier labs are realizing that software/models themselves don’t have real moats and move to embodied ai.

SOTA 80% means a practically useless robot. What are they really imagining their ICP to be here?


What does this comment mean?


It’s unclear to me what their desired outcome for a blog post like this. If you’ve ever worked in a robotics setting, 80% implies that 20% of your autonomous actions are incorrect. Imagine if this were the case for autonomous driving where your car misbehaves 1 in every 5 actions it takes.

Posts like this just reminds me of the end to end demos AV companies built in the early days using a single camera - only to realize that it’s harder than it looks years later into development.


The ICP question was more around the model itself. Are they looking to license it to robotics companies? Do they imagine that devs at robotics companies would be willing to deploy these models as a black box?


Exactly! We experimented with a whole bunch of video encoding techniques for LLMs here: https://vlm-run.github.io/mm/encoders/#video


Pretty terribly expensive way to watch a video with Claude.

Use Gemini or some local VLM to do this way more efficiently. We spent quite a bit of time on video understanding, and Claude will just burn tokens.

Check out this library: https://vlm-run.github.io/mm/

You can swap models and try out different encoding methods for videos (https://vlm-run.github.io/mm/encoders/#video)


Exactly this. Gemini is best at this. Just give it video link - YouTube works best - and it will analyse the video.


Really, does this work now? What about NotebookLM? I was using it a lot until i realised it was only analysing the transcripts and not the video because i was mostly using it for technical ones with important charts.


NotebookLM still uses the transcript method I think. But Gemini is wonderful. I have been using it to analyze the youtube videos of wrestling matches (trying to build a fan website for WXM, the best pro wrestling promotion to come out of India in a while). It does move by move analysis, audience reaction based match flow tracking, isolates interesting parts of the video (big moves, botches, story beats etc). I have run some experiments to get video editing plans out of it. I think I can combine it with something like remotion skill to make highlight videos.

Edit: BTW, you can analyze about 8 hours a day on free tier.


NotebookLM was transcript-only when I last checked. Gemini proper does ingest video natively (samples ~1fps server-side). This tool is for everything that can't — Claude, ChatGPT web, local models — it turns the video into frames + transcript on your machine so any of them can read it.


It can tell you what’s on the screen at given point in time. My pipeline is mostly around simple questions like “does this video contain cars?” Not sure if it can spot charts on screen.


mm can spot charts on a screen. did you try it?


Seems cool from the docs page, I was about to give it a shot but https://github.com/vlm-run/mm goes 404 …


It’s unclear if that’s intentional since it’s listed also under open source on the main company site: https://www.vlm.run/open-source/mm


We were planning to open-source this soon, but jumped the gun and posted about the video encoders here since it seemed relevant.

In either case, here you go, it's public now: https://github.com/vlm-run/mm.


Do you mean that Gemini is most token-efficent at watching videos? Is that the case for e.g. just giving it a video in the browser? I admit, I dont give LLMs videos as I just assume it'll burn too many tokens.


Yes, Gemini is very token efficient at video. It also has "lower resolution" options which can make it even cheaper if. With Gemini 3.1 flash lite an hour of video works out to $0.24 at the API rates.


Assuming that's your project, the GitHub link from the PyPi page is a 404.



Awesome!


This is neat. I'd love to figure out a sequence of emojis that triggers the LLM in ways that puzzles a human.


That’s a pretty large binary for simply loading images.

In all honesty, opencv has stood the test of time and I’m certain newer LLMs will likely not attempt to rewrite it from scratch.

P.S. I’ve been a user since the IplImage days, circa 2007, and I’d still consider using it over most CV libraries today.


> I’m certain newer LLMs will likely not attempt to rewrite it from scratch

Sooner or later a Rust developer will try.


You can build it yourself and end up with a much smaller binary (and many more optimisations).


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: