Oct 15 8 days ago 07:50 Consolidate conclusions into a dedicated section; rewrite the Q&A/correction trail as a coherent cost breakdown; report the compute gap as 1.6-1.8x over two reruns 07:43 Correct the steady-state paging figure (1.06 MiB/tok, not 14.8); document the no-preload design and the remaining 1.62x compute-kernel gap at >=16 GiB 07:42 Correct the steady-state paging figure (1.06 MiB/tok, not 14.8) and document why there is no preload path plus the remaining 1.62x compute-kernel gap 07:38 Explain why 16/32 GiB still read SSD: it is a one-time 14.4 GiB cold fill, misses freeze at 2421/2496 07:19 Add model card metadata (base_model, tags, license) 07:18 EPQ expert paging for Xing4.0-29B-A4B IQ4_NL: strict RAM matrix, I/O audit, correctness suite 07:18 initial commit 
