I'm not very good with hardware compared to most enthusiasts and was hoping /maho/ could please give me some help with something of a new build. I've been offered a free PC upgrade through my mother's company but I'm not confident in what I'm doing. From what I can read at the moment, AMD has overtaken Nvidia for practical affordability and price:performance ratio, but doesn't offer very high end GPUs. I've also been recommended an X3D CPU by an acquaintance, but I don't believe Intel offers an equivilent yet.

I've been using Intel+Nvidia all my life just because it's what I had help with from my older cousin growing up. My main use case is games, as well as some Adobe suite work mostly Lightroom and Photoshop and a small amount of video editing. I don't do anything in the vein of AI, and don't really play many new release games; my current hardware is able to run most of what I play without issue with the most demanding thing being Baldur's Gate 3, heavily modded Skyrim, or Team Fortress 2. Although TF2 is an older game it struggles on my current hardware and I believe simply upgrading the CPU will do wonders for it. I intend to play TES6 when it eventually releases.

https://nz.pcpartpicker.com/list/vgZ9Lc

I am already using the case, CPU fan and RAM in the above link. From what I can understand the 9800X3D and 9950X3D are functionally the same for my use case with the latter just being better for workstation use. According to this website and this review

https://www.videocardbenchmark.net/video_lookup.php?gpu=Radeon+RX+9070+XT&id=5956
https://gamersnexus.net/gpus/amd-radeon-rx-9070-xt-gpu-review-benchmarks-vs-5070-ti-5070-7900-xt-sapphire-pulse

The 5070 Ti is slightly better than the 9070 XT at a slightly higher price point, but is overpriced compared to RRP.
One of the difficulties in doing this is that everything I get has to be from a local store called PBTech due to use of a business account to recieve a discount, so something like the 7900 XT is not an option. I'd like to please ask for help choosing a graphics card, motherboard and PSU, then RAM if necessary.

>>3937
>cgpeers? Can you see if something I've been looking for is there?
Sorry I didn't see that post until now because someone bumped the thread. kissu has been too fast and active!
Yes I can, what are you looking for, person two weeks in the past?

>>3938
VRAM is the biggest thing for AI. Sure, the 5090 and 4090 will have some speed over the 3090 but having the 24gb of VRAM on the 3090 is what allows you to do many things at all. A 3090 is like having a $400,000 Ferrari instead of a $500,000 Lamborghini. It's not an exaggeration that the change from a model being offloaded some to RAM from VRAM is like 30-50x slower, if it's possible at all.

>so I'm checking for motherboards with two PCIe 5.0 slots

Be aware that the newer cards are huge. I can't physically fit my old card in the motherboard on the second slot. I wonder if someone will make an aftermarket cooler that's thin, but I bet it will be expensive as all of these things are.

>For the CPU, I'm currently looking at the i9-12900KS,
When I did research for CPU stuff AMD was the kind of mixed work/gaming CPUs and I went with the 9800XCD which had trouble staying in stock.
Now to read the new post...

>>3940
even more important
¥always call out a greener when it's andy

>>4109
If you have no AI plans, AMD is definitely something you can do as a cheaper and much better price point. The 9950X3D is probably something you should only buy if you already have specific plans to make use of it extensely, which most people don't. The 9800 is already really powerful at its price and that's why it's so popular.

>One of the difficulties in doing this is that everything I get has to be from a local store ...I'd like to please ask for help choosing a graphics card, motherboard and PSU, then RAM if necessary.

I'm not sure how we could help since we can't go to the store with you. Maybe you can go there and take notes or something?

>>4112
Sorry, to be clear I more meant I can't order off Newegg, Amazon etc or buy second hand because the purchase is being made by my mother's company and has to be done locally. It's a chain and they have a website: https://www.pbtech.co.nz/
Another store that sometimes has more niche products is https://computerlounge.co.nz/. I don't know if I can buy from them in this instance however.
The 9800 sounds good. I'm sometimes a bit wary of buying a very popular part, or even something popular outside of hardware, because it often seems like it's for a common use case, but I suppose I fit such a case when it comes to PC use. And thank you for reading my post, sorry I got a bit rambly.

>>4113
Hmm... Well, I'm not too familiar with AMD cards but I'd say you want at LEAST 12gb of VRAM. Personally I don't think you should bother buying a card with less than 16gb. If you can't afford the extra $100 New Zealand bucks (or the company rather) then save until you can.
But... hmm.. I just really don't know AMD stuff. I haven't had an AMD GPU since like 2002 or something. I'm just too paranoid over driver issues with obscure emulators and stuff. It's better these days, though.

>>2269
get a radeon for the same chipset at the same price chances are it's gonna be much better

oh wait january post. how are you doing anon? enjoying your new card?

>>4109
both 5070 and 9070xt are such a weird things. they are 5 years more recent then radeon 6800xt and still losing to it in many categories.

actually, speaking of weird, lets start how fucking bizzare it is that 5060 and 5070 are base 8 gb vram when 3060 had 12 lol

>>4196
>how are you doing anon? enjoying your new card?
Doing good! With AI I'm using my graphics card far, far more than I did when they were just for gaming. It used to be I'd get a decent card and then use it like 20 days of the year with specific games, but now I'm using it practically every day. I made a new SDXL NoobAI super merge to meet my demands and it took far less time since I was able to generate the images in a fraction of the time. I'm currently trying to hook up SillyTavern to SD so I can automatically generate images for scenes while the text AI is offloaded to OpenRouter. AI Video is kind of fun, but it doesn't hold the 'style' so it's not as good as I would hope. Unfortunately training video LORAs seems outside the realm even of the 5090 so I'm not sure when that would improve. People generally seem focused on porn, which is all well and good, but it quickly reverts to a generic appearance which kills the purpose of having 2D AI at all.

The only problem is AI makes it hard to find motivation to exist in real life or to create things, but I'm trying. The good news is that this card allows me to have blender open doing simple stuff unrelated to rendering, so I can AI prompt in the background while working towards other goals... slowly.

>>4215
what is the advantage of having your own ai model over using something like novelai or even chatgpt?

>>4216
Well, local models are better than online for the fact that you can customize them by adding LORAs and stuff, most local people do that. ChatGPT is of no interest to me because it's censored corporate stuff I can't use for escapism.
But when it comes to my obsessive model merging...
The model I made is a merge of like 8 other NoobAI-based merges. The advantage is selecting things I like and trying to optimize things. Some models are great at "creativity" (random mutations/errors) while others are good at poses or backgrounds or tag recognition and so on. By spending a bunch of time selectively merging different layers to try and isolate the best parts of each I try to maximize everything without sacrificing other things. For example layer OUT00 here has a huge influence on body types, so maybe someone else's model does great here, but layer OUT01 which controls the general coloring and shading is terrible, so I can take Layer OUT01 from the former and inject it into another model without the terrible colors.

>>4217
err I mean OUT00 is injected. Whatever.

>>4217
interesting. I use AI for human-level analysis. I mostly just use notebooklm, and feed it people's books, chatlogs, personal correspondence and any other source I can find, and use AI to draw this person's psychological profile, or look for inconsistencies in their narratives etc. Do you think I could benefit from having my own model?

>>4220
You're talking about text LLMs so the investment is like $10,000 to get a computer server rack setup thing that has like 800gb of fast RAM to load the entirety of the deepseek model, which has been the premiere local model for 2025 so far. Most people doing online stuff use the same model, but hosted by a company that largely charges nothing to use it. (I think they log the prompts for future training data)
The online-only models like GPT and Claude boast far larger context windows and tool use/agents so unfortunately the online stuff has a noticeable gulf in usability.

>>4221
Oh, but note that you can of course load smaller and weaker text models, but it's very much a hobby thing. People don't want to use something more prone to hallucinating when it comes to programming help and the roleplaying is far more dry since the data isn't there.

AI Image Gen: Conquered. Well, hands are still an issue, but not as much. I'd wager I have the best 2D model in the known world for my specific needs.
AI Text Gen: Relatively great, I'm not rich so I can't do the best online models. I'm incorporating image gen and video emotion avatars.
AI Captioning: Yep. Got the NSFW tag thing going on.
AI Video: Making progress, but some weird errors keep popping up.

The 5090 has been very useful.
How can I make money from this while still being lazy...

>>4244
>How can I make money from this while still being lazy...
There’s some really obvious and cynical answers, but you probably don’t want to do them.

>>4246
Yeah, probably. Stuff like spamming social media or actual art sites with an AI deluge until I get enough attention?
I don't want to make the world worse, I have my basic needs met so I don't need this to survive. Probably best to keep it as a tool to help me masturbate and design characters/scenarios for games.

Sadly this low sampler setup for video doesn't actually seem to work for me and I need to do the long generation times.
It seems like you can offload some VRAM blocks to CPU (or whatever the terminology) was, but since I can barely squeak by and fit this specific resolution inside 32gb of VRAM at 12 second run time, my generation time is pretty amazing. 5 minutes for this. It would probably be about 15 hours on the 3080 if I had, if I could do it at all.

I am now in the exact same situation and feel like my mind has already surrendered to making the worst financial decision of my life. I was originally gonna replace my 960 with a 3090 Ti, but now that I have seen just how much quicker the 5090 is for AI video alone, a feature I could actually really use, I know I wouldn't feel completely content with going for something lesser.
After doing some research, I could theoretically get one and pay in monthly installments of $100 at no interest, which coincidentally is the max. amount I usually save in a month, but I'd need to save like $50 on top of that for next year when my usual expenses increase in cost. I really need to find some loaded wallets on my weekly trip to the store..

>>4436
what do you need AI for? if you just play games, and used to be content with ancient shit like a 960
just buy radeon 6800xt or 7900xtx

Stay away from overpriced nvidia crap

>>4437
Not the same Anonymous but how is Radeon for AI? Been thinking of getting AMD GPUs for LLM/image/video gen use but most of youtube says AMD ones aren't as good as nvidia ones.

>>4438
AMD didn't try competing in that area until they released R9700 a week ago. Currently it's still not available for purchase. It is inferior to 5090, but ought to be cheaper.
Rather then watching youtube, use sites like nanoreview or technicalcity to compare the specs and look at benchmarks yourself. youtubers are paid shills.

>>4436
OP here. I can say video was fun to mess with, but the time involved makes it more annoying than fun. Maybe it's just me, though. Video stuff is clearly made with realism in mind as it often struggles very much to maintain a 2D style if you tell it to do too many different things. Loops are the way to go since it's restricted to matching the end frame to the starting frame, reducing the freedom it has. If you're someone that finds 8 second loops of video far more erotic than still images then, yes, it's worth it. For me, though, it's usually too uncanny.
ACTUALLY NEVERMIND!
I forgot that a few weeks ago I spent like 20 hours messing with the ComfyUI nodes to combine two different workflows to try and make some good 2D looping. This actually looks pretty good. Well, definitely problems with her looking non-Kuon-y for a few frames, but that's why you'd generate like 20 videos and then go do something else. Alas, the randomness of AI is still here.

People keep making advances on making LORAs and similar things that allow you to generate faster, but... wait for it... it's made for real images.
Video models also use 'natural' language instead of booru tags, which I find really annoying, too. You can at least use other AI models to write the prompts or make suggestions. Video is definitely frontier territory, so keep that in mind. Also I want to say again how much I hate comfyUI.

For multitasking it's truly amazing. I can keep SD loaded and still play a game like Palworld and not feel anything, I can even have SD generating some images at the same time! Not video, though, video uses all the VRAM. If you want to do AI RP stuff with an on line text model you could also have some better stuff loaded to go along with it, like local voice model and stuff. All in all, the more you use AI the more you'll use the card.
It's kind of weird, now that graphic cards are also used for AI I use them about 5000% more often and now I actually feel like the cost is justified in a way.

>>4438
If you're going to be doing AI stuff you need nvidia. There's no way around it. nvidia has CUDA, AMD has... nothing. nvidia's value wouldn't have become so extremely high if it wasn't a monopoly.

>>4438
AMD GPUs are much better if you're actually interested in driving a display(s) and having support 10 years from now. You don't have to rely on AMD to build drivers against the latest kernel and it is much less of a headache to keep it going.

Nvidia GPUs are effectively a black box. You're much better off buying them used and running them as second GPU. But even then you might not be able to continue using them forever. The only Nividia GPU I own now is not really that old but they no longer build drivers against recent kernels. So I'm stuck either running an old kernel or not using it at all. I hate having to plan my entire OS around the GPU.

I refuse to support them any longer because I've gotten burned too many times. There is really no point in using AI locally either. It isn't nearly as impressive as everyone is pretending it is.

>>4442
It's good to have principles and stuff, but the question is whether or not AMD is good for AI. The answer is a resounding 'no'. There isn't a choice. You pay the monopoly its entry fee or you're forced to sit out.

>>4440
I’m fascinated by how you’re doing this. What’s your workflow like when making these generations? Is it at all hard to make it similar to what you’re picturing in your head?

>>4442
using AI locally is great if you have actually a reason to use it
if all you use it on is generating pictures of waifus and gifs like the nvidia shill does then yeah, it's just a luxury for the rich

also another argument: price needs to be considered. if you are serious about AI, you will need to invest. ideally at least 128GB of ram and multiple gpus. Here's the question of how much cheaper the R9700 will be once they start selling is separately. Projections are up to half the price of 5090. So here's the real question: what is better, a single 5090 or two R7900s? The answer to this question will be known once benchmarks start appearing. I'd wait a few weeks with buying a 5090 until then.

>>4449
>or two R7900

typo btw, meant two R9700s of course

>>4446
Holy crap someone else is interested in this amazing technology?!
I only mess with image-to-video model as I can't imagine these are any good at generating 2D stuff from nothing. It's not going to know characters and stuff, so having a starting image is a huge requirement.
My basic process is:

1. Generate a good source image in SD
2. Use AI to do a basic natural language observation of the image + make suggestions on what to animate. I'm not sure if this is doing anything, I need to test it more. For example is "She moves her left hand smoothly to pick up the glass" really any better than "hand moves with glass"? I'm guessing it isn't.
3. Generate the video with a workflow thingie that took me a few days to actually understand.

Just like with image gen there's LORAs, but they're quite a hassle. Most of them are geared for realism as I said earlier so it's not that great, but some are really really nice for NSFW stuff.

>>4449
>the nvidia shill
How much do you think nvidia is paying me to call their GPUs ridiculously expensive and to take months beating myself up over wanting to buy one?

You've entered my thread to call me a shill because I'm genuinely excited about a technology and sharing my personal experience on it to advise others. Not only that, but you're doing it purely out of ignorance.

Stable Diffusion/Image Generation? CUDA.
Audio Generation? CUDA.
Image Captioning? CUDA... I think? (most likely)
Video Generation? It took me a solid week to get the latest CUDA drivers to work with sage attention and pytorch.
Local Text Generation? This one is actually less dependant on CUDA. You can run text gen on CPU, but there's a large speed penalty. This is how people are running deepseek locally without spending $80,000.

Everything is built on CUDA and CUDA is nvidia.
Since you're proclaiming AMD over and over, do you know the name of AMD's CUDA analog?
I'll give you a second to think about it.


It's ZLUDA. It's not their own thing. It's a hack to try and get nvidia's CUDA to run on an AMD GPU and people have been working on it for years at this point. AMD as a company is not working towards it, just random people online. AMD forfeited.
If you think I have loyalty towards a monopolistic shit company like nvidia you are sorely mistaken. I would love it if there was actual competition in AI, but that is not our present world.

>what is better, a single 5090 or two R7900s?
For AI, the 5090. Unfortunately it's not even a comparison. AMD prices similarly to nvidia to such an extent that people said AMD blew it this year with its high prices, choosing to nickel and dime people instead of expanding their market share. They joined nvidia with the price gouging. Their "AI" card is being sold at a fraction of the nvidia price because it will be a fraction as useful.

>>4454
>You've entered my thread to call me a shill because
>my thread

no such thing, we are anonymous and this is a public place

>>4437
Aside from the obvious NSFW use, it has gigantic potential for 2D editing, think Eroge-style Motion Graphics that don't really follow a red line in terms of design. That's on top of other smaller time-saving benefits that add up over time, also related to workflow. I actually don't really play a lot of games myself.

>>4440
Thank you for your input, I was planning on asking!
What are your observations regarding the progression / optimization of local generation, does the community make an effort to try and make models less resource-intensive over time? Or is it similar to unoptimized AAA games with huge raw assets no one cares to compress, instead expecting the consumer to upgrade early? Because that's my main worry here. As someone who doesn't use camera footage, I know I will be set for non-AI editing for a decade+, but I haven't really been able to get a feeling regarding AI. For example, does image generation nowadays use less VRAM at more speed and equal / good results on, say, 20-series cards?

Also, what's up with the supposed lack of 2D-focused models? I think to remember that the threads on 4/h/ (might've been /trash/ instead, it's been a while) had something relatively exciting to report about every few days. To be fair, that was when the AI craze was still new and it was only related to images.

>>4456
>does the community make an effort to try and make models less resource-intensive over time?

Yes and no. Models themselves can't be made more efficient, it's basically a settled technology once it gets to the hobbyist's hands. It's the corporations and research teams that could make more efficient models, and then some rich guy could rent a data center to create a 2D model based on that tech. But he's not making a new technology, just training one based on another.

There have been a few efficiency improvements over the years in regards to VRAM usage and stuff, but I don't think much of it is new new, it's a few years old now. There's probably some gains for image gen dependant on ComfyUI, but I HATE ComfyUI and do everything I can outside of it. I know for video there's a VRAM offloader thing in ComfyUI that increases generation time but allows you to offload some stuff to RAM so you can generate the thing at all. I use it even though I have 32gb of VRAM. Video is THAT demanding once you get to certain sizes or lengths.

Anyway, those VRAM gains for image gen are offset by the newer models being made to generate larger images, which requires more VRAM. People also want to use the newer bells and whistles like upscaling and adetailer and controlnet and other stuff which each have their own VRAM cost. But you can definitely enjoy image gen with SDXL with 8GB or possibly even less if you heavily restrict active programs and limit yourself to maybe 1-2 images generated at once. I still have hardware acceleration disabled in all my browsers to free up a few hundred megs of VRAM.

There have been more improvements to LORAs/models that can generate things in fewer steps, which means less generation time which is most noticeable on people with weaker cards. However it generally comes at a cost somewhere, usually quality, and I never liked the ones I tried. (and the ones I didn't try require manual installation by altering code and that's above my brain). For video models, these faster generation things come at a cost not just of quality but they also tend to heavily restrict the animation in various ways, so each one I've tried left me disappointed. Except "lightx2" LORA, that one seems good and it's in all of my recent video things.

>Also, what's up with the supposed lack of 2D-focused models?
There's no money in it I would assume, or at least comparably. For video it's probably just that RL video data is far easier to come by when it comes to data scraping. And if somehow your dataset lacks something, grab a smartphone and record it yourself.

>>4457
Man, how disappointing about the excessive demands for video.. even if your 5 minute result from above does look very promising. What do you assume a 10s video would take to generate if you go for the maximum resolution that still fits into your VRAM with no offloading?

I stopped experimenting around the time people talked about LORAs, I just used the so-called "NAI leak" back then and decided that waiting 5 minutes for a single super low resolution image with horrible quality isn't worth the hassle. I've always wanted to get into training models with my favorite artists, but that's completely out of the question with my current card. I shall see if it's really gonna be the 5090, might just be the best "budget pick" on Nvidia when it comes to VRAM, though I haven't checked past workstation cards yet.

Either way, thank you very much for all the info! I shall make great use of it in the future.

>>4458
>Man, how disappointing about the excessive demands for video..
That's just for the extremes. 720x960 is quite huge for a video size and then the length greatly increases the VRAM usage. I can do 81 frames at 720x960 at like 22ish GB of VRAM.
My 'block offset' is 9 blocks (moved to RAM) when I do 161 frames, which is certainly on the extreme end of things. Videos tend to get errors with the colors at this length so I don't often try it. It might be a workflow thing, though.

>What do you assume a 10s video would take to generate
It goes by frames and not seconds, with adjustable framerate. 12 or so FPS seems to be about the slowest I can go without consciously noticing it's slow. 81 frames seems to be sweet spot, possibly what the model was trained at, so it would be a bit under 7 seconds.

>>4458
>What do you assume a 10s video would take to generate if you go for the maximum resolution that still fits into your VRAM with no offloading?

Oh and I'll time and test this later, but can't right now. It should be like 5-10 minutes when I push it to the limit of frames.

>>4460
Okay, I did some testing. I'm using 12 frames a second.

720x960 resolution and 81 frames: 25gb of VRAM used and it completed in 171 seconds. I forgot that NOT offload blocks to RAM, but I assume the gains wouldn't be that major. Maybe a dozen seconds?
Don't know why it's behaving so weird with the colors. Such is video AI.

>>4477
161 frames. 9 blocks offloaded to VRAM. 29gb used.
317 seconds. Maybe it's the pure white BG that's making it so weird.

>>4459
>>4460
>>4477
>>4478
Thanks so much again! It's nice to see how the first example seems to try being more creative with the animation. Most of the videos I have seen have this "animated PNG" feeling to them, similar to Gacha games. Have you read anything about RAM speed making a significant difference when offloading?

Also, would you mind sharing your sources when it comes to AI news that revolve around models designed for us 2D-focused people? Or really anything that would still benefit that genre.
I sometimes stumble upon random YT videos that showcase a large variety of things that seem incredibly interesting (and increase my 5090 itch), but it's just so much at once.

Last thing: We need more Koupen!

>>4477
>>4478
Bleh, I now notice mistypes. I mean it's loaded from VRAM and not to it. The "offload blocks" thing moves it to RAM or something to reduce VRAM usage.

>>4481

>Also, would you mind sharing your sources when it comes to AI news that revolve around models designed for us 2D-focused people?
I just check civit.ai now and then to look at models directly. AI news, particularly youtube, is full of scammers and opportunists that promise you the moon so you'll click their thumbnail and are the same scum that did (and sometimes still do) the crypto hype stuff. I'll also rarely scan AI generals since it's easy to find informative posts as the guy will usually type more than 8 words, which makes it stand out from 95% of posts inside a general thread.

>>3893
>I feel immense guilt and shame over it
As you should! For ~$6K you can now get a DGX-1 system with 256GB of pooled VRAM. If you had saved your money instead of buying a 4090 and then a 5090 we could have an actual Kissu AI server. Tsk tsk tsk.

https://www.ebay.com/itm/167512048877

>>4762
I figure this is how AI stuff will go if you want to be serious about it. Buying a GPU and building a desktop around it isn't the best idea. If you get an Nvidia GPU now you'll have to deal with them dropping driver support in 5-10 years. This really sucks on Linux because you end up stuck on an old kernel or running the terrible FOSS drivers.

You're better off buying a GPU that provides full driver source code for your display devices and building a server for working with AI stuff. It's designed for that anyway. You send data off to be processed and it comes back.

Either that or wait until they start providing dedicated AI chips. I figure that is coming down the pike in the next few years. Built right into the motherboard (or at least a socket for it). I see us going to unified RAM soon instead of having the current RAM/VRAM split. Maybe this will improve the situation with buying a card to run displays and real time graphics. Right now things are so bad I've put off buying a new GPU for almost a decade. I can't justify spending what they're asking when my old one still does what I need.

I've moved most of the things I used to do on my desktop off to servers. All my files are stored there and it's running several things like a music+video daemon that can serve the data with metadata to whatever other device I want to use. Over the LAN and the internet. It has been much cheaper in the long run and having access to everything when I'm away from my main desktop has been very nice. It reduces how much I spend on other things like laptop and small handheld devices.

Prices are going up on the stuff you need for that to. But things like enterprise switches+routers+servers with server grade CPUs+RAM+HDDs are very cheap if you shop around. They dump so much of that on to the used market every year. The only thing that's really expensive still is the stuff you need to run a fiber network inside of your home instead of on CAT5/6 cabling. Which can push 10Gbps anyway and will probably get another speed increase with time.

>>4762
It does seem cool, possibly, but I'm beating myself up after scraping together $2800 so spending $6000 is simply impossible.
I'm really not sure how or if that thing you linked could be used for my AI purposes. From what I've seen having two GPUs instead of one complicates things, it's not just a plug and play kind of deal. You load THIS into GPU A, and THAT into GPU B, but you need to merge the result HERE and so on. I'm not some machine learning guy, I just run what other people have made.

The crux of the matter, though, is that this stuff is only as good as its adopters.
Right now there's been good progress in local video stuff because more people are getting 24gb and higher VRAM, which allows the hobbyists to create and share things. Without a community of people with the hardware it's not that useful.
As for text LLM stuff, 256GB is unfortunately still too low to load the best local models. There's a massive gulf and mid-range stuff largely doesn't exist.

>>4771
>still too low to load the best local models
256GB should be enough to run GLM-4.5-Air at FP16 (225GB), or GLM-4.5 at around INT4 (~190GB; Full-size is 755GB at FP16). Air has rough parity with GPT4o-mini, and GLM-4.5 rough parity with GPT4o. That's purely for knowledge-level tasks, though. No idea whether they're any good for RP or whatnot.

>>4919
Well, that's good to hear. That's still extremely out of my price range, though. Maybe we can start a Kissu Fund where everyone gives me thousands of dollars. What's the context window with that kind of setup?
That sounds really, really nice. Well, I guess it depends on the (E)RP potential, though. Ultimately people want the best of the best for things that matter, like programming, but ERP can be 'eh, good enough" if you're in the mood.

>>4992
>What's the context window with that kind of setup?
Probably in the range of 5-15K tokens... So... Usable with aggressive summarization, but maybe not so great for long, detail-oriented back and forths. Something something.... you can quantize the LLM, but not the context window (KV cache)?

Anyways, you can actually use GLM-4.5-Air and GLM-4.5 on https://chat.z.ai/. It's a bit faster than Deepseek, but nowhere near as fast as ChatGPT.

>>4994
Are you still sending me the thousands of dollars?

>>4995
Sure, I'll just need the money upfront and I'll 10x whatever you give me.