Pull requests / #624

#624 bench: community report — RTX 4090, Ryzen 9 7950X, Flash-Next IQ3_S

closed · @Dmitry-B · 0 comentarios · En GitHub

BenchmarksNVIDIA / CUDAModels & quants

Descripción

Two paired runs at 4K/32K/128K prompt tokens with three repetitions each, TTFT via streaming, memory peaks, and six needle recall checks (6/6 found): one with Russian prompts, one matching the community code-explanation style for direct comparison with the RTX 5090 report. Engine 0.1.38, context 143,360, draft_vocab=cyrillic.

En el sitio

Enlaces a install, modelos, releases.