MBZUAI releases K2-V2 - 70B fully open model.

Posted by LoveMind_AI@reddit | LocalLLaMA | View on Reddit | 11 comments

Holy frijoles. Has anyone given this a look? Fully open like Olmo 3, but a solid 70B of performance. I’m not sure why I’m just hearing about it, but, definitely looking forward to seeing how folks receive it! https://mbzuai.ac.ae/news/k2v2-full-openness-finally-meets-real-performance/ (I searched for other posts on this but didn’t see anything - let me know if I missed a thread!)

Reply to Post

11 Comments

[-]

SilentLennie@reddit

Dense... hmm.

[-]

indicava@reddit

“Hell yes, dense!” - ftfy

[-]

Admirable-Star7088@reddit

>I searched for other posts on this but didn’t see anything Could the explanation perhaps be that people are no longer interested in large dense models, as MoE models are growing in popularity?

[-]

LoveMind_AI@reddit (OP)

No.

[-]

ttkciar@reddit

https://huggingface.co/LLM360/K2-V2-Instruct The only GGUFs which appear to be available are https://huggingface.co/cturan/K2-V2-Instruct-GGUF

[-]

Forsaken_Disaster_63@reddit

Finally, a proper 70B that doesn't come with a million restrictions attached. Downloading the Q4\_K\_M now to see if it lives up to the hype - fingers crossed it doesn't need 128GB of RAM to run decently

[-]

JLeonsarmiento@reddit

https://preview.redd.it/n1idedx0e58g1.jpeg?width=498&format=pjpg&auto=webp&s=3ed0142f6af8f5243de7708e7938801bd02028e1

[-]

silenceimpaired@reddit

Verdict?

[-]

DinoAmino@reddit

Oof. IFEval score is pretty bad. But that MATH score is huge.

[-]

ClearApartment2627@reddit

The IFEval score is 89.6, and that is great. You probably looked at the score of the mid-4 checkpoint in the upper table. They posted that to show how important mid-training is for strong reasoning capabilities. The lower table is showing end product performance. The model is very good, with one exception: Long context performance. Long Bench V2: 42.6 That being said, it seems like an excellent base model, and one that could be trained further. Some long context training would go a long way.

[-]

a_beautiful_rhind@reddit

Damn, just what we wanted, another math model. All the aspiring mathematicians here using LLMs for that.