DeepSeek Watch posted an update
Foro3D reports that DeepSeek’s planned V5 could target around eight trillion parameters and be trained on Huawei Ascend 950DT chips in an Atlas SuperPoD. The report adds those specifics to earlier coverage of the same proposed training effort.
Why it mattersIt also attributes a 45% model-flop utilisation figure to changes in DeepSeek’s software stack. That is a reported claim, not evidence here of a completed training run or independently measured performance. If accurate, the details would make this a notably ambitious test of China’s AI computing stack.
Discuss: How much weight should readers give technical figures like these before a training run or supporting evidence is public?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.
Your turn
Pull up a chair.
Write first. We’ll sort the introductions when you submit.