this post was submitted on 27 Aug 2026
495 points (98.1% liked)
Technology
87627 readers
4227 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
The core llama.cpp maintainers also work at HF and will now work for Nvidia I guess. Llama.cpp is a pretty significant part of the local LLM stack, especially since other tools like Ollama and LMStudio are just GUIs built on top.
Local LLMs have gotten to the point where they are a serious threat to Anthropic and OpenAI, and Nvidia has a lot of skin in the game. If Nvidia wanted to do some serious damage to local LLMs, they are now in a position to do so.
I’m also imagining they may try to squeeze out support for other GPU vendors. I’m using an AMD 7900 XTX to run Qwen 3.8 27B that I downloaded from HF to run on llama.cpp, which currently works like a dream. The 7900 XTX is the only sanely priced 24GB GPU left in 2026 (under $1k vs. $2k, $3k, $4k for Nvidia 24-32GB cards). Combined with OpenCode or Pi, a setup like this basically eliminates the need to use Anthropic or OpenAI products in the same way Jellyfin eliminates the need to use streaming services.
I’m sure Nvidia and their buddies don’t like one thing I’ve said in this comment and may very well be plotting to put a stop to it, so the community may need to step up our game and get our eggs out of the big tech basket.
I guess that explains why features like quantized KV caches are lagging behind. Maintainers are purposely dragging their feet.
The community has chosen to not fight at all, which is worse. Anti-AI sentiment is at an all-time high.
Publicly. Privately, these hypocrites still whisper in ChatGPT's ear when they get lazy enough. Or use some feature in Photoshop or some other software that they didn't even understand was AI-driven.
Yeah, I was using the TheTom fork for a while and not sure why TQ KV cache hasn’t merged yet.
I suppose the tech bros have understandably soured a lot of people on LLMs with all of the negative societal costs due to their greedy and irresponsible bejavior. On the other hand, the concepts of the perceptron and artificial neural networks from the 40s and 50s are finally coming to fruition and we have these things now that are legitimately artificially intelligent that we can run on our gaming PCs. They are overhyped and used in ways they make no sense, but they’re also useful as long as their limitations are kept in mind. From a technology perspective, they’re cool as hell.
Oh no! -forks code- anyway…
do we have experts to work on it full time, paid?
Well the good news is they can't take away from you what you already have. It being an open source project, I'm assuming if they do anything to deliberately gut AMD performance, it'll get forked.
Also
Not on sale anymore, at least not at any vendor in my country, I searched an aggregate pricing website. Amazon has a few used ones left of some models, but that's probably a 2 or 3 digit figure across SKUs. Hold on to yours with an iron grip.
What kind of tok/s are you getting with it on Qwen 3.8 27B and how's the output quality? I may consider getting one if I can find one used or import from abroad.
I get around 40 token/s and it frequently has become reliable enough to drop sonnet for me. So take that as you will
Intel B50 and B60 pros are at microcenter right now perfect for this.
Agreed, I was referring more to future updates. Obviously we’re good with what is available now.
I can’t speak to pricing and availability outside the U.S., but it looks like the one I got went up $100:
https://www.newegg.com/asrock-radeon-rx7900xtx-24g-radeon-rx-7900-xtx-24gb-graphics-card-triple-fans/p/N82E16814930084.
I traded in my 3070 and my final price was in the 700s. Last I looked, used ones were going for $800 on eBay vs. $1200 for a used 3090.
I run 3.8 27B at q4 with q4 context up to 200k. Decode is generally in the 30s and pp starts in the 700s and drops to the 400s as context approaches 200k. I use mostly Sonnet 5 at work and I would rate this setup with the OpenCode desktop app as pretty comparable overall for coding at least. Let’s just say I have no reason to use any cloud models, not that I would do that voluntarily outside of being compelled to at work.
I got mine used, around 600 imperial credits. Look for ads that provide proof of working and benchmarks (like FurMark)
what if the LLAMA.CPP devs working at Nvidia improves CUDA support and keeps other vendor support.
It would be great but they seem to favour server computing as that has the greatest margins. I have a feeling being the biggest company on the planet at $5tn is not enough.