BigMoeOnEdge: streaming MoE inference on Android, built on llama.cpp (ggml_org).
DeepSeek V4 Flash IQ2_M (92 GB) running on a 12 GB midrange phone at ~1 tok/s. Qwen3.6-35B-A3B up to 7-8 tok/s, Gemma 26B and Qwen3-30B in the same tok/s range. CPU only. Experts stream from UFS
Xianyu is the largest AI hotspot in China and the most profitable self-media platform in China🙃 The previous one is MiniMax H3
这次是DeepSeek Harness Ps: 现在还有人蹲Mac mini M4 吗？🙂
DeepSeek Harness finished the same coding task almost 3× faster than Claude Code.
But Claude still produced the better result. Julian Goldie and Kasra gave both systems the same one-shot prompt: build an animated accounting website and a working Tetris game. The results: 00:41
Trying the Deepseek Harness Beta again was unexpected
The Herness agent created by Deepseek, there are no privileged cores that have to be patched. Everything (adapter model, tool registry, session log, even the loop agent itself) can be changed from the configuration. Main capabilities:
You have reached the end of the archive
All of deepseek