Tencent yuanbao has pulled ahead of bytedance ai apps.
this is going to be interesting to watch
Artificial Intelligence thread
- Thread starter 9dashline
- Start date
Tencent yuanbao has pulled ahead of bytedance ai apps.
this is going to be interesting to watch
"China's smartphone maker Honor pledged $10 billion in AI investments over the next five years and announced on Sunday a deepening partnership with Google.
The investment plan, revealed at Mobile World Congress (MWC) in Barcelona, is designed to reposition the firm from a smartphone player into an "AI device ecosystem company".
Honor meanwhile displayed their new deepfake detection technology on its Magic 7 Pro smartphone, a world's first.
The new Hornor also features a new AI agent, is expected to be rolled out globally later this year.
Honor is somewhat of an upstart in the smartphone world, after spinning off from Huawei in 2020 when the Chinese tech giant was hit with US sanctions. Since then, Honor has looked to expand outside of China and push into the higher-end part of the market where Apple and Samsung play.[...]"
The investment plan, revealed at Mobile World Congress (MWC) in Barcelona, is designed to reposition the firm from a smartphone player into an "AI device ecosystem company".
Honor meanwhile displayed their new deepfake detection technology on its Magic 7 Pro smartphone, a world's first.
The new Hornor also features a new AI agent, is expected to be rolled out globally later this year.
Honor is somewhat of an upstart in the smartphone world, after spinning off from Huawei in 2020 when the Chinese tech giant was hit with US sanctions. Since then, Honor has looked to expand outside of China and push into the higher-end part of the market where Apple and Samsung play.[...]"
Bytedance has open-sourced an optimization system called Comet, which improves the execution efficiency of the Mixed Experts Model (MoE) by overlapping fine-grained computation and communication.
It seems similar to the DualPipe & DeepEP open-sourced by DeepSeek, but not as complicated. The advantage is that it can be used with only a few lines of code changes. The experimental results of Bytedance are that Comet speeds up the execution of a single MoE layer by 1.96 times and end-to-end execution by 1.71 times. It has been deployed in large-scale GPU clusters, saving millions of GPU hours
Paper: arxiv.org/pdf/2502.19811
Code: github.com/bytedance/flux
It seems similar to the DualPipe & DeepEP open-sourced by DeepSeek, but not as complicated. The advantage is that it can be used with only a few lines of code changes. The experimental results of Bytedance are that Comet speeds up the execution of a single MoE layer by 1.96 times and end-to-end execution by 1.71 times. It has been deployed in large-scale GPU clusters, saving millions of GPU hours
Paper: arxiv.org/pdf/2502.19811
Code: github.com/bytedance/flux
Last edited:
QwQ 32B is out! Comparable benchmark to DeepSeek R1 with about 20x less parameters!
kind of expected imo. more time and effort in RL learning, the models keep getting a lot better.
DeepSeek was smart to be the first open source reasoning model, that gave it a lot of oxygen that other models can't match.
Now, just waiting for their next model to come out.
Manus is out and trending on Weibo. Apparently, better than Deep Research based on GAIA benchmark.
I can see many people using this.
Kling 1.6
Tencent's new Hunyuan AI model for image to video
looks like this is the one
Tencent's new Hunyuan AI model for image to video
looks like this is the one
Day0 supports Moore's Thread and quickly supports Tongyi Qianwen QwQ-32B open source model
On March 6, the Alibaba Cloud team officially open-sourced a new inference model, Tongyi Qianwen QwQ-32B. Moore Threads quickly and efficiently completed support for Qianwen QwQ-32B within 2 hours after the release of the model. QwQ-32B, which is deployed based on the large language model high-speed inference framework vLLM and the MT Transformer inference engine, has demonstrated excellent inference performance and stability in actual operation, fully demonstrating the strong advantages of the MUSA architecture and full-featured GPU in ecological compatibility and fast support.
Moore Threads has opened this achievement to the model square "Kuae Factory". Kuae Factory is a model display center created by Moore Threads with great effort, aiming to provide users with model capability experience based on Moore Threads KUAE (KUAE) intelligent computing cluster support. Users can experience the powerful inference model performance and innovative technology of QwQ-32B by visiting Kuae Factory or clicking "Read original text".
this guy seems to kind of legit based on the stuff he posted and his audience size.
Basically saying that much of Nvidia sales in Singapore, 100B over 2 years will come to China. Since the Chinese big tech can't buy the smuggled chips directly, they get installed in the public data centers and then rented out to the major cloud/tech providers like Alibaba, Tencent and ByteDance.
Apparently, China's major data center build up are in Gansu (this was the one I posted earlier in the year that said it will be 100 EFLOPS by end of 2025) and Shanghai. My guess is that those are still under estimates.
since Lenovo and Meta-X partnered up for domestic DeepSeek all in one machine, it has sold over 1000 units covering almost 10000 cards, Covering medical, education and manufacturing.
N260 can infer 14B model at 110-130% of main line GPU.
C500 build AI server can run full blooded version at 3725.1 tps
Saudi Aramco praising its integration of DeepSeek in its Aramco data center.