Tag
The author optimized DeepSeek V4.1 Flash for Apple M3 Ultra, achieving up to 40 t/s decode speed with DSpark speculative decoding while maintaining byte-identical accuracy to the upstream model.
Optimizations for GLM 5.3 Flash on Apple M3 Ultra achieve up to 550 t/s prefill and 38 t/s inference speed through kernel fusion and efficient memory use, without quality loss.
An author improved MLX performance on an M3 Ultra Studio to achieve 73 tok/s for a 4-bit Qwen3.8-Flash-Next model, which shows intelligence scores matching GPT-5.6 Sol, highlighting the growing potential of local AI.
The article details progress in achieving over 50 tokens per second inference speed for DeepSeek-V4-Flash on M3 Ultra using the DwarfStar inference engine, with improvements in performance and quality metrics.
The author achieved a 12x speedup for DeepSeek V4 Flash on a Mac Studio M3 Ultra by optimizing kernels and implementing effective caching strategies, reducing chat turn latency from 6-20 seconds to 1.6 seconds.
An uncensored version of the GLM5.2 754B parameter model (231GB GGUF) was successfully deployed on a Mac Studio M3 Ultra with 512GB RAM, achieving approximately 3.6 tokens/s.
A user reports that running Hermes Agent as a game master using DeepSeek V4 Flash locally on M3 Ultra yields nearly identical quality to the online version.
AWS has secured a large number of Apple's M3 Ultra Mac Studio units for cloud services, while regular consumers face continued shortages and limited availability.
Antirez reports that DeepSeek v4 PRO runs well on a Mac Studio M3 Ultra with 512GB RAM using 2-bit quantization, achieving 130 t/s prefill and 13 t/s generation.
A researcher shares their home compute setup for MLX and AI research, featuring M3 Ultra with 512GB, RTX PRO 6000 with 96GB, and M3 Max with 96GB for model porting and stress testing.