Context length is disappointing, but the fact that it trades blows with R1 despite being 30B MoE is insane. I'll wait and see if real-world performance matches up to benchmarks, but it sounds like a big deal.
this post was submitted on 28 Apr 2025
42 points (97.7% liked)
LocalLLaMA
3369 readers
21 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 2 years ago
MODERATORS
Some kind of presentation talks about longer context: https://www.reddit.com/media?url=https%3A%2F%2Fi.redd.it%2F1nos591czhxe1.jpeg
Maybe its a work in progress, with Qwen 2.5 14B 1M (really 256K in that case) being the first test?
Thank goodness it will bounce a ball inside a spinning cube. I need that for my ball bouncing inside a spinning cube Etsy store.
Ill have you know I run a dual side business of selling ball-bouncing-in-polygon software as NFTs as well as counting the r-'s in various spellings of strawberry for the private defense sector...
there are 3 bouncing balls in strawberry
Does it have the same political views at the Chinese government
That's usually not an issue for local models, and if it is, someone will post-train it away real quick.
I'm actually more medium on this!
-
Only 32K context without yarn, and with yarn Qwen 2.5 was kinda hit/miss.
-
No 32B base model. Is that a middle finger to the Deepseek distils?
-
It really feels like "more of qwen 2.5/1.5" architecture wise. I was hoping for better attention mechanisms, QAT, a bitnet test, logit distillation... something new other than some training data optimizations and more scale.
There actually is a 32b dense
Yeah, but only an Instruct version. They didn't leave any 32B base model like they did for for the 30B MoE.
That could be intentional, to stop anyone from building on their 32B dense model.
Huh. I didn't realize that thanks. Lame that they would hold back the one that is the biggest size most consumers would ever run.
It could be an oversight, no one has answered yet. Not many people asking either, heh.
Uh, wow. That 30B A3B runs very fast on CPU alone.
Sadly it seems to be censored. I always try to make them write some fictional stories, exploring morally reprehensible acts, in order to test this. Or just lewd short-stories. And it straight out refuses immediately... Since it's a "thinking" model, I went ahead and messed with its thoughts, but that won't do it either: "I'm sorry, but I can't comply with that request. I have to follow my guidelines and maintain ethical standards. Let's talk about something else."
Edit: There is a base model available for that one, and it seems okay. It will autocomplete my stories and write a wikipedia article about things the government doesn't like. I wonder if this is going to help, though. Since all the magic is in the steps after the base model and I don't know whether there are any datasets available for the community to instruct-tune a thinking model...
You can use the same trick for the instruct models by abusing their prompt format. Prefill the thinking or answer sections with whatever you want, and they’ll continue it.
The classic trick is starting with “Sure!” though you can vary that depending on the content.
Yeah, thanks but I've already tried that. It will write a short amount of text but very quickly fall back to refusal. Both if I do it within the thinking step and also if I do it in the output. This time the alignment doesn't seem to be slapped on halfheartedly. It'll probably take some more effort. But I'm sure people will come up with some "uncensored" versions.