Artificial Intelligence thread

O

OptimusLion

Junior Member
Registered Member
Tongyi Qianwen has open-sourced a new visual understanding model, Qwen2.5-VL, the flagship visual language model of the Qwen model family, with three sizes: 3B, 7B, and 72B.

The main features of Qwen2.5-VL:

◆Visual understanding: Qwen2.5-VL is not only good at identifying common objects such as flowers, birds, fish, and insects, but also can analyze text, charts, icons, graphics, and layouts in images.
◆Agent: Qwen2.5-VL directly acts as a visual agent, can reason and use tools dynamically, and has the initial ability to use computers and mobile phones.
◆Understanding long videos and capturing events: Qwen2.5-VL can understand videos of more than 1 hour, and this time it has the new ability to capture events by accurately locating related video clips.
◆Visual positioning: Qwen2.5-VL can accurately locate objects in images by generating bounding boxes or points, and can provide stable JSON output for coordinates and attributes.
◆Structured output: For invoices, forms, tables and other data, Qwen2.5-VL supports structured output of its content, which benefits applications in finance, commerce and other fields.

Please, Log in or Register to view URLs content!
 
00CuriousObserver

00CuriousObserver

Junior Member
Registered Member
Alright. So I might have some clues on how the "DeepSeek has 50k H100s" came about

Just sharing some info with a theory .


In November last year, a semiconductor analysis company called SemiAnalysis stated that DeepSeek possesses over 50,000 Hopper GPUs.

Please, Log in or Register to view URLs content!
(Dylan is the boss of SemiAnalysis)

This SemiAnalysis is reportedly claimed to have "multiple leading enterprises using their data"

Please, Log in or Register to view URLs content!


However, Hopper is an architecture. Hopper GPUs include not only the H100 but also models like the A100, H20, and H800, which are officially exported to China. At the same time, they also claimed that DeepSeek has H100s, suggesting they could circumvent sanctions.

Please, Log in or Register to view URLs content!


This news circulated within the industry. When it reached Alexandr Wang during an interview with CNBC, whether intentionally or unintentionally, the reference to "50,000 Hopper GPUs, including some H100s" became "50,000 H100 GPUs."

Please, Log in or Register to view URLs content!


This version then spread widely, despite Dylan repeatedly debunking the rumors and stating that "some client executives have misunderstandings."

Please, Log in or Register to view URLs content!


Please, Log in or Register to view URLs content!



Also, DeepSeek possessing 50,000 units of a certain model GPU does not contradict what is written in their paper. Having 50,000 units does not imply they have used or need to use all 50,000 units
 
Last edited:
tphuang

tphuang

General
Staff member
Super Moderator
VIP Professional
Registered Member
I don't know what is up with people just treating these threads that should be high quality tech thread as their meme discussion playground. If you can't be serious about some of these discussions, they will get removed.
 
tphuang

tphuang

General
Staff member
Super Moderator
VIP Professional
Registered Member
https://twitter.com/i/web/status/1884224743366078963

Perplexity is now integrated with DeepSeek R1. Although I have not checked the pricing on this. Here is the thing though, there is noting preventing 3rd party from taking their own local build of R1 and using it or creating apps with it. It's super easy. The data security argument doesn't work when you can host these things entirely outside of China.
 
You must log in or register to reply here.
Top