Kapitsa 0.2.26 adds SSD offloading and Apple Silicon benchmarks
Kapitsa contributor Ye announced version 0.2.26, adding support for DiffuSpace's Clover 35B. He reported that Qwen Flash-Next 180B reads prompts 46% faster and Qwen3.8 27B writes 8% faster than its previous M5 Max results. The release also adds SSD offloading for Flash-Next 180B: contributor Ye said experts that do not fit in RAM can stream from an SSD, enabling a reported 61 tokens per second on a 48 GB MacBook Pro. Benchmarks on a 2026 M5 Max MacBook Pro with 128 GB of memory covered six models at context lengths from 1K to 32K. The figures came from a project contributor, who invited others to submit comparable results and warned that the new codebase could contain prefill or decode bugs.
The release expands local-model support on Apple Silicon and allows Flash-Next 180B to run when the model cannot fit fully in RAM. Its performance claims come from contributor benchmarks and remain open to further community validation.
#Kapitsa#KFlash#Qwen Flash-Next 180B#Qwen3.8 27B#Clover 35B#Qwen3.6 35B-A3B#Flint 40B#Nemotron Bolt 74B#DiffuSpace