For the best experience on desktop, install the Chrome extension to track your reading on news.ycombinator.com
Hacker Newsnew | past | comments | ask | show | jobs | submit | history | mattnewton's commentsregister

If China keeps releasing LLMs with such permissive licenses it probably works better since the pretraining rnd is subsidized and de-risked - but that’s a big if.

The exception being the token embeddings and lm head (which scale with the number of tokens the model knows and presumably you need a smaller number in the tokenizer for only English and python). But those are a pretty small % of the total model weights on most LLM sizes

Honestly the 27b dense one punches way above its weight in a lot of domains, especially coding in my testing, so I think you will probably be disappointed.

There was a 3.5 122B 10A release -

https://huggingface.co/Qwen/Qwen3.5-122B-A10B


I tried it for a bit, and It was not really worth its size. It got swept up in all the other AI news recently, but laguna s 2.1 I think is the best ~100B moe model right now

I didn't mention it above, but Laguna S is my other favorite model. I use Qwen a lot more, it's smaller and faster, but I like to switch to Laguna when I feel like I need a "heavy hitter" for certain huge or complex tasks.

What on earth hardwares do you guys have to be able to run 100gb models locally?! That's crazy! I'm here struggling to even get 27b models to run in somewhat usable way

Haha I'm on an Mac Studio with an M1 Ultra, 64gb ram. I bought it when it first came out, it just happens to be good for local LLMs. I have to use a smaller quant of Laguna S though (I think 4-bit? Not at my machine to check), as 8-bit and full size definitely don't fit in the 64gb I have.

brb, going to see if 2nd hand mac studios are available!

Yeah, a good rule of thumb is that the weights take up ~100% of the size of the model, so 100B bytes (8-bit quant) would be, well, 100GB and a 4-bit quant would be half that.

Oh, that is a useful rule to know! Thanks!

That's not exactly the math. Theres also vram needed for context. I operate several 72-128 GB machines and the larger the context the slower they go.

And the context takes space +kv cache. KV cache drives usefulness as your context grows, it needs to pull the kv cache.

Simplified, the context has to be run on every turn, so the KV cache supplies the processed tokens, so it just needs the new inpute.


I don't know if that's, well, a rule of thumb, it might be, well, straight multiplication.

Well, yeah, the straight multiplication, er, well, "follows" from the rule of thumb. Hope that helps!

Saying, well a rule of thumb is, well, 100 billion bytes is a 100 gigabytes, is well, not a rule of thumb. It is, well, just the common definition.

Right. The rule of thumb is that the overhead size of the model that's not the weights is so vastly outweighed by the actual number of weights that it can be disregarded. My shorthand for that was to write "the weights take up ~100% of the size of the model". What then "follows", both in the sense that the explanation is written after the rule as well as that it logically follows, is that, well, 100 billion bytes is, you know, 100 GB. I don't see why you're, like, paying so much attention to this?

Right. That's, well, not a rule of thumb. Asking how much of a bottle of water is, well, water, and someone says "well, a good rule of thumb is that it's all water", is just an answer. You don't need an estimate when, well, there is nothing to estimate.

Yeah, well, that's just, like, your opinion, man.

Anyway, more seriously, I hope it's obvious by now that I don't particularly care that my means of communication is so offensive to you. I think you should, like, cry a river, build a bridge, then, well... get over it, you know?

And on the, er, "topic"? A rule of thumb for me does not have to be one for you, even if it's explicitly presented as such a rule. I thought that would be obvious but, well, here we are. Anyway, how's life been treating you?


Strix Halo, 128GB RAM. I got a refurbished Corsair AI Workstation for a smoking price ($2100) about two months ago. Lucky timing that it was in stock.

Strix Halo as well. Bought it for $1,800 new on sale and shoved an extra 4tb drive into it. Been amazing for local AI. Maybe not the absolute fastest thing (usually around 30t/s depending on the task) but has been awesome for a local AI box that I can solar power.

Nice. Mind sharing the solar side of your setup?

Couple of rack mount batteries and roughly 5kw of solar panels. Feeds into a subpanel so I can flip it when I want a couple rooms of solar on the house, or hook a generator up if needed. Can't power the entire house, but works well for thinks like computers, lighting, etc. And if I want to expand, just throw on more panels, or realistically, just throw on more batteries to store the juice.

Thanks - that’s what I want to do!

Yeah, here I am sitting deeply deeply deeply regretting not buying couple CMP 170HX at $200 or $350, knowing I could just flip them ethically at purchase price if nothing came of it... I could have just casually built a 128GB dual A100 local AI monster

I'm working with a lab that has a few Ampere GPUs on infiniband and they are just not compatible with the latest quants and vLLM updates. FP8 is about as low as you can go.

But they're reportedly a soft nerfed GA100 64GB/40GB at $1200, that's not more expensive and certainly can't be slower than a Mac Studio.

quantized + offload

I have an RX 6700 XT with 12gb vram and 64gb system ram. running dense models like 27b is difficult, but i can run IQ4/IQ5 qwen 122b-a10b or 35b-a3b at ~20tok/s


IQ4 qwen 122b-a10b would mean 61GB total size and 5GB active, so about 5GB of the model loaded into GPURAM plus any generated context, and 61GB of weights loaded into system RAM? I don't know if that math is correct, but does that run well? Wouldn't that only leave 3GB of system RAM?

Usually theyre quantized. Also, there was a window where AMD 395+ W/128GB was just a high end $2500 hardware with unified gpu memory.

dgx spark, nvfp4 so I have spare room for KV cache (context)

MoE models can use system memory along with a GPU.

and get high token bandwidth?

Not badly so because MoE models(identifiable by "CoolName-xxxB-AxxB" naming scheme) have bunch of branches in the middle that only one out of all gets non-zero values. Each of branches aka "Experts" as well as top/bottom parts are significantly smaller than the whole, and so CPU emulation of CUDA operations mixed with GPU taking as much as possible become not so out of question, unlike for dense models("CoolName-xxxB" without "-AxxB")

Similar to a spark, which isn't blazing fast but usable.

Using qwen 3.6 27b for local coding as well and downloaded Laguna s 2.1 but haven't had time to give it a full spin yet.

Curious for any more experiences


I agree. 27b dense really did seem like the sweet spot.

Because demand for inference tokens is above supply

This was a big concern for earlier models, but with modern CoT trained models they should be able to come to the conclusion entirely in the thinking trace.


I think a) the labs are releasing very fast and b) why would they implement the long tail of app features when they can effectively sell tokens to every user to write their own version of the app, which is what is currently happening?


Because, as the gp pointed out, if the cost is least to the labs, then why not reap the benefits too?

Hypothetical. Assume you can in fact point agents at a tool and say "replicate it. Make no mistakes". You then have software being instantly copy-able.

Assume these agents can then be pointed to a customer feedback board in perpetuity and they autonomously upgrade the software over time. They analyze usage patterns and behave like PMs figuring out what to prune and what to build. Then the maintenance part of the stack also goes to zero.

Over time, the highest margin competitiveness will go to the distributor of the tokens. Aka the AI model makers.

In a world like that (which the frontier labs claim is within a year or two of happening) it feels like it's only a matter of time before they opt to own the entire stack down to the consumer apps. Kind of like Amazon deciding they want to knock off products doing well and then favour their own product over the original seller.

My guess is that if the capability arrives the only reason the frontier labs don't move to own the entire stack immediately is because of optics. Boil the frog instead.


There is more to selling software than writing it. You have market, support, and sell. Do you think their resources are well spent doing that across the gamut of software? Of course not; companies specialize.


Very similar to how cloud providers love renting servers to you to run your bank or software business; instead of running these businesses themselves.


Isn't the promise that LLM can do all this better than any human? Or at least in few short months? Surely marketing, support and selling is just case of right prompt?


Why do your think Meta doesn’t make money? Their ad platforms are incredibly lucrative.


There are plenty of services to send mail form the internet for a small fee, so this will only discourage the most poor candidates and add friction for the best ones.


The "small fee" is the point though as it stops practically all automated submissions. I doubt the price of a stamp excludes poor candidates unless you think that poor people could never get jobs before online applications were a thing (and internet connections and phones cost money as well).


It’s not that poor people can’t afford a stamp it’s that they aren’t going to spend money on an automated service or stamp if there are other places to apply to that don’t require this.

So I think it’s going to lower recall a a lot - it reduces volume of good and bad actors equally. Anecdotally, I’m not going to bother with such an application unless it’s my top choice or I’m desperate. For small companies that probably means throwing the babies out with the bathwater.


I don't know what the options are these days but MCI used to do it. Yeah, it's some friction on both sides.


Because PDFs are a nightmare of a format and the only thing that’s is reasonably guaranteed about them is they will render to an image that people can read, the parsing of which will be much less token efficient than the equivalent text


I agree with you, but every non-engineer I know using these tools 100% will drag and drop a PDF into a chatbot. Anthropic and OpenAI as companies who are selling their products to all sorts of businesses should have a much better means of handling this nightmare of a format because it is so pervasive and so obviously what so many of their customers are going to drop into the product.


Why would they spend a ton of effort ensuring that their customers spend less money on them?

Token economics also are weird. If you design a fancy new frontend that for example uses a cheap model to parse a PDF into text that is fed into an expensive model, you will probably spend more money because you are on API payscale rather than the "max plan" payscale.


I think they’ve just decided that vision gives the best results and the token issue will take care of itself.


> I agree with you, but every non-engineer I know using these tools 100% will drag and drop a PDF into a chatbot

I'm an engineer and use my coding agent to deal with PDFs all the time. It can reach for unix tools if it needs them.

I don't think I understand why this is a problem - it uses tokens, but it removes drudgery. This is the entire promise of the technology.


I’m saying there is basically no way to both make vlms able to understand the long tail of PDFs where the layout conveys information (like charts and tables) and to make it as token efficient as text formats. Current approaches have mostly chosen to work more often than not at the cost of token efficiency.


Once you pay full price for tokens plus margin, it is better for the company to burn as many tokens as possible.

For the same reason as why the oil companies want everyone to use large cars.


For anyone needing to do this, the answer is to convert it to an image first. Far smaller, LLMs work well with them (even in some pretty insane use cases I've seen), and, along with human review, it can be a huge productivity gain that results in structured data.


I agree with your recomendation, but converting a pdf to an image is by no means smaller. PDFs are much closer to SVGs then to jpegs.


Why can't I just take a screenshot of the PDF and feed that into the llm?


Since I'm almost never interested in the formatting I run all pdf files through pdftotext from the Poppler library before llm use.


Been building various LLM+PDF pipelines at work. As soon as you need to e.g. parse tables etc. it becomes a lot of hard work!


Consider applying for YC's Fall 2026 batch! Applications are open till July 27.

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

Search:

HN For You