MODEL / COMMUNITY REAP BUILD
Qwen3.8 Flash Next REAP-288
Community-pruned conversion, not the stock checkpoint. The 39 GB model / ~40.57 GB total resident path requires oMLX 0.6.4+ and streams only the sparse PLE table from NVMe. A published M4 Pro 48 GB test measured 26–29 tok/s and recommended 7,680 tokens; do not generalize it to arbitrary models or Macs.
Fonte primáriaParâmetros totais176B
Memória estimada~40.57 GBREAP-288 · MLX 4-bit · PLE on NVMe
Contexto7.6K recommended on measured 48 GB MacText
LicenseQwen CommunityMultilingual; post-pruning quality not independently established
01 / IDEAL PARA
- 48 GB Mac experiments
- coding agents
- model-specific NVMe streaming
02 / AJUSTE ESTIMADO
03 / RODAR COM
oMLXAdvanced
omlx serve --model-dir ~/models --memory-guard-gb 48Atenção
Community-pruned conversion, not the stock checkpoint. The 39 GB model / ~40.57 GB total resident path requires oMLX 0.6.4+ and streams only the sparse PLE table from NVMe. A published M4 Pro 48 GB test measured 26–29 tok/s and recommended 7,680 tokens; do not generalize it to arbitrary models or Macs.