[Top topic overseas] The "free agent environment" that brings together Kimi K3 and DeepSeek is a hot topic!
! The operation that links the NVIDIA Build prototyping frame (Free Endpoint) and the OSS execution frame "DeepSeek Harness (dsh)" and switches multiple models from the browser UI at no cost is attracting attention. ・npx deepseek-ai/dsh web
Three free AI writers worth knowing
Kimi K3 ranks above GPT-5.6 Sol on the current EQ-Bench creative-writing leaderboard, DeepSeek V4 offers a huge context window, and Gemma 4 runs locally. Rankings and free plans can change. If you want me to help automate manual work in your
This is a video I posted on Station B explaining the DeepSeek Harness architecture.
The animation parts inside were all done in one go using Opus5.5. It’s so cool! Come watch and experience what AI that understands videos means. It’s not necessarily the coolest, but the resulting image really fits what I want to express. In contrast, I used GPT and Gemini to make one version each, which can only be called web PPT.
DeepSeek V4.1 Flash (2-bit MoE, experts streamed from SSD) on M5 Max
Decode 12.7 -> 24.1 tok/s averaged over 2048 tokens (~1.9x), output bit-identical to upstream main. Video: a 256-token run. ~18 tok/s while the expert cache warms up, then 24-26 tok/s. What this branch changed:
You have reached the end of the archive
All of deepseek