Intriguing. Should be interesting to compare this to Qwen3.8 (27B). I've been plenty impressed with Qwen's ability, but boy does it like to overthink things, even after getting it off the default xhigh reasoning, it'll take 20 minutes to plan something, thinking so much it'll have to compress its context multiple times to reach a single answer. The quality of the answer is great, but there has to be a middle ground.
this post was submitted on 04 Sep 2026
60 points (96.9% liked)
LocalLLaMA
5120 readers
51 users here now
Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.
Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.
As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.
Rules:
Rule 1 - No harassment or personal character attacks of community members. I.E no namecalling, no generalizing entire groups of people that make up our community, no baseless personal insults.
Rule 2 - No comparing artificial intelligence/machine learning models to cryptocurrency. I.E no comparing the usefulness of models to that of NFTs, no comparing the resource usage required to train a model is anything close to maintaining a blockchain/ mining for crypto, no implying its just a fad/bubble that will leave people with nothing of value when it burst.
Rule 3 - No comparing artificial intelligence/machine learning to simple text prediction algorithms. I.E statements such as "llms are basically just simple text predictions like what your phone keyboard autocorrect uses, and they're still using the same algorithms since <over 10 years ago>.
Rule 4 - No implying that models are devoid of purpose or potential for enriching peoples lives.
founded 3 years ago
MODERATORS
36B-A4B!
Very interested in this
Is this the first actually open source model?
There is the Swiss project Apertus AI, and I think I've read about others but don't remember the names.
That being said, if the benchmarks translate to real utility, it is the first competitive open source model, I'd say.
We disclose how the data was constructed and mixed when redistribution is not possible.
So sounds like some data is unreleased, I wonder how much.
My guess is they don't really have the rights to release data copyrighted by others.
We are releasing intermediate checkpoints, training data or detailed data-construction recipes
👀
Their new attention mechanism sounds super interesting.
Can't wait for it to hit ollama so I can easily add it into my flow. I really need a better ollama alternative where I can easily download, manage, and run gguf locally and tie it in with opencode, zed, and open-webui
A word of advice, consider switching to something else.
This was the article that had me looking for alternatives but nothing seems to have the same tight integration that ollama has with the 3 tools I use the most
Does llama.cpp router mode or llama-swap not meet your requirements? What doesn't work?
As far as I can tell, open-webui doesn't have a way to manage gguf models via llama.cpp / llama-swap. It can leverage the models just not a great experience managing (downloading, deleting, etc). I can definitely adjust my workflow to use another tool, it just means an additional hurdle compared to my current flow but it sounds like it's worth it from a performance standpoint
Eh, are you really using that many new models where sshing into your server / cding into your model folder, wgetting a url and configuring your interfaces every now and then is that much of an issue? I can't imagine using more than like one or two new models a month unless there's some insane string of releases or something.
Also, I mean, everyone's setup is different, but there's a significant amount of performance you're potentially leaving on the table by not using llama.cpp, potentially in the double-digit percentages. (Plus, if you have a fairly recent Nvidia setup and are willing to wait a bit for the latest models, ik_llama.cpp is a fantastic fork that I've found can get way better performance on most models than even llama.cpp.)
Llama.cpp of course! After I did the switch, never looked back. Add llama-swap in front of you switch between models often.
llama-server has router mode built-in; you don't need separate software to swap models. I keep a list of my models in a presets.ini file.
Yes but llama server does not unload models to fit a different model. At least, didn't managed to have just llama.cpp swap between models that fill up my vram... The second one would fail load.
Ah. I POST to '/models/unload' when I switch models as part of my tooling -- if llama-swap is doing that for you, then I get why you'd want it.
Yes, llama swap also give you a pretty web based statistics of all the calls and runs for every model with t/s and more statistics.
It's pretty neat... You can also load and unload models manually, define groups for models that fit together in vram and so on.
It gives you that automation that people coming from ollama are used to.
under the hood all it does is running llama serve. You convert your models.ini to a yaml file 1:1 (plus a few more flexibility).
This is the way 👆