DeepSeek V4.1F Q4 on M3 Ultra with native DSpark MTP (40tps / 800tps)

A user optimized DeepSeek V4.1 Flash for the M3 Ultra, building on previous GLM optimizations. This involved forking antirez/ds4 to improve performance, particularly for agent models. Key improvements include decode speeds increasing from 16.6 to 31.3 t/s for 8k context and 14.0 to 28.3 t/s for 300k context. DSpark performance also saw gains, with agent turns improving from 31.3 to 41.3 t/s, making the model more practical for real-time use.