Is it time to short NVIDIA again? DeepSeekV3 0324

Posted by Sicarius_The_First@reddit | LocalLLaMA | View on Reddit | 10 comments

Will llama-4 be again delayed? Will we see massive shorts on the stock market? The horror and terror big tech is experiencing right now, I can only imagine. What do you think? How good is it gonna be? What are the implications? [https://huggingface.co/deepseek-ai/DeepSeek-V3-0324](https://huggingface.co/deepseek-ai/DeepSeek-V3-0324)

Reply to Post

10 Comments

[-]

anon235340346823@reddit

Minor upgrade according to DeepSeek. https://preview.redd.it/an8t15httnqe1.png?width=1294&format=png&auto=webp&s=42c777b1be96278fcf61e011bba283cfec8dcf54

[-]

EtadanikM@reddit

No bench marks released probably means it isn’t world beating. Hence not going to V4 as well. When Deep Seek drops something big you’ll know it because they’ll also drop the bench marks & the research report.

[-]

Sicarius_The_First@reddit (OP)

I wonder why they released it then, it cost a lot to train such a big model, even if you're "the chinese"

[-]

AppearanceHeavy6724@reddit

To probably collect the feedback I guess. Keep in mind their primary target is not Westerners but China. They are collecting the feedback first and foremost from Chinese.

[-]

EtadanikM@reddit

Same reason Anthropic dropped Claude 3.6 without even renaming 3.5 (hence the version jump later to 3.7). It’s an improvement but isn’t worth bragging about it because it isn’t revolutionary.

[-]

kellencs@reddit

[-]

Sicarius_The_First@reddit (OP)

I just assume Zack will not release LLAMA-4 just to be DOA Unlike the Chinese models, Meta spends BILLIONS, they have share holders to give answers to.

[-]

BillyWillyNillyTimmy@reddit

Considering that Digits is shaping up to be a let-down, with a worse memory bandwidth than a GTX 1080 (but fingers crossed CUDA will help), Nvidia might actually not rebound on AI.

[-]

Sicarius_The_First@reddit (OP)

TBH the new blackwell lineup for workstation is actually REALLY GOOD, and it's even not expensive at all for what you get. The problem is... availability. Probably no one (except big companies) will get it for MSRP.

[-]

Sicarius_The_First@reddit (OP)

Just an example, the 48GB card costs pretty much the same as an old Ampere A6000 on ebay right now. https://preview.redd.it/crwtrl11rnqe1.png?width=835&format=png&auto=webp&s=cccef30920d2a84d1621edba732e35eda6bfd9e9