DeepSeek

Monday 16 November

Sunday 15 November

Baseten publishes DeepSeek engine, reports 2,060 tokens per second on eight B300s

AI-generated summary Baseten published an inference engine for the DeepSeek V4.1 Flash model, reporting a speed of 2,060 output tokens per second on eight Nvidia B300 GPUs. Teams with compatible hardware can run the DeepSeek weights on the engine, which was written in Rust and CUDA and released under the Elastic License 2.0.

Earlier:

DeepSeek V3 architecture implemented in 190 lines of JAX

AI-generated summary Vaibhaav Tiwari has released an open-source implementation of the DeepSeek V3 architecture using 190 lines of JAX code. The project was trained on a MacBook Air using public-domain detective fiction books and runs without heavy frameworks.

Also covered byTheorem WeeklyEl Solitario
Earlier:

Saturday 14 November

Researchers allege DeepSeek, Kimi, and other Chinese AI models were trained using banned Nvidia Blackwell GPUs — Researchers use new GPU 'fingerprinting' technique to identify the GPUs used to train open models

AI-generated summary Researchers at UnemploymentLabs allege that Nvidia Blackwell chips were used to train Chinese artificial intelligence models including DeepSeek V4 and Kimi K3.1. The researchers applied a fingerprinting technique analyzing weight initialization patterns to identify the hardware families used for training open-weight models.

Also covered byThe RegisterTopFiveState of AI
Earlier: