I was able to visualize DeepSeek-V4-Flash inference using DGX Spark x2.
Including MoE's 43 layer 256 (11,008) expert usage, MTP draft acceptance/rejection status and RDMA communication throughput between two devices.
Power Dynamics CEO jenzhuscott explains why NVIDIA buying Hugging Face could be a brilliant vertical integration play
"This is a very smart acquisition if it goes through because right now, closed AI labs like Anthropic and OpenAI, they all started moving to hardware. DeepSeek
DeepSeek raised the rate of its API service with differentiated prices by peak and off-peak hours, the latter by half.
The scheme is in effect from August 17, from 9 a.m. to 12 p.m. and from 2 p.m. to 6 p.m. OpenAI and Anthropic cut fees to defend their share. 🇨🇳🤖📊
Testing Qwen 3.8 Flash Next Q4 on the DGX Spark: Day 34.
Asked it to create Pac-Man, and after 24 minutes, it failed. DeepSeek is the only model that's come close. Interesting to see how future versions progress!
You have reached the end of the archive
All of deepseek