Kapitsa 0.2.26 Released: Benchmarks Multiple Models on Apple Silicon

Open-source inference engine Kapitsa released version 0.2.26, optimized for Apple Silicon with custom Metal kernels. On M5 Max, Qwen3.6 35B-A3B achieves 346 tok/s decode and 6,036 tok/s prefill; Flash-Next 180B runs on a 48GB MacBook Pro at 61 tok/s via SSD offloading.

• Alibaba quietly announces Qwen 42026-11-06
• User praises OrcaCyber-Nano-1.0 fine-tune2026-11-06
• NVIDIA engineer promotes open models via YouTube2026-11-05
• Box taps Nvidia's Nemotron for agent-first content retrieval system2026-11-05
• User Alleges Qwen3.8-27B Quietly Weakened2026-11-05