holy cow, an 8b comparing to o3-mini
View original threadholy cow, an 8b comparing to o3-mini
@sungkim.bsky.social
DeepSeek's DeepSeek-R1-0528-Qwen3-8B
They distilled the chain-of-thought from DeepSeek-R1-0528 to post-train Qwen3 8B Base, obtaining DeepSeek-R1-0528-Qwen3-8B.
huggingface.co/deepseek-ai/...
They distilled the chain-of-thought from DeepSeek-R1-0528 to post-train Qwen3 8B Base, obtaining DeepSeek-R1-0528-Qwen3-8B.
huggingface.co/deepseek-ai/...
36
2
1 hour later
i grabbed the biblically correct bf16 version of R1-0529-8b off unsloth and... oh my dog this shit is slow
apparently ollama is upcasting it to fp32
apparently ollama is upcasting it to fp32
6
first impressions: poor adherence
this AI is avoiding my real question and instead throwing into question my competence for even asking such a question
this AI is avoiding my real question and instead throwing into question my competence for even asking such a question
5
uh.. thanks?
6
safe to say it doesn't quite feel like the full size version bsky.app/profile/timk...
@timkellogg.me
i’m noticing that R1-0528 doesn’t use words like “Wait”, it just keeps digging deeper on its own, more naturally
3
9 hours later
I gave it this problem and I think I've hit an infinite loop
2
i'm also starting to see weird behavior. here — it adds ** around two lines. in other spots it just has random ** that aren't part of formatting. Granted, it's the thought trace, but it feels like it's lost it's mind (tbf i would too if i was forced to solve this)
2
let it run?
1