Drilldown: Recipe

Qwen3.8 27B on 1x 8 GB (W4A16 vLLM) (1) · Qwen3.8 Flash Next on 1x 8 GB (ElectroHobo 3 bpw) (1) · Qwen3.8 Flash Next on 1x 8 GB (iIIusi0n SSD PLE) (1) · Qwen3.8 Flash Next on 2x 8 GB (DUFRelic PP2) (1)

Qwen3.8-27B (dense) (1) · Qwen3.8-Flash-Next (2) · Qwen3.8-Flash-Next (125B MoE per the recipe) (1)

AutoRound 3 bpw + MTP: klee100/Qwen3.8-Flash-Next-AutoRound-3bpw-MTP; 95.37 GiB BF16 PLE table on SSD (1) · AutoRound mixed 2/3-bit (klee100, 2.998 effective bpw, 47 GiB on card) plus halt95's FP8 n-gram table in host RAM (1) · W4A16 (INT4 weights, Marlin kernels) (1) · W4A16 INT4 g128 with the PLE table repacked INT4 g32: zankich/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV; KV BF16 (1)

Morrowmake/vllm-cmp170hx image with patch files mounted read-only (1) · vLLM (1) · vLLM a5a30471 plus a 13-file downstream patch (native Linux AIO PLE reader) (1) · vLLM v0.30.0 + zankich INT4-PLE plugin + six hash-guarded patch layers (1)

None (1) · 0.29.1rc1.dev402+ga5a30471f.ple1 (1) · recipe 0.2.1; engine 18305948ac (1) · vLLM v0.30.0 (ced6857a); plugin zankich/vllm v0.29.0z (6609bfef) (1)

8 GB · Other · None

1 (3) · 2 (1)

single card · engine-managed expert tiering across VRAM, host RAM and NVMe · engine-managed split with expert tiering across VRAM, RAM and NVMe · layer split (no tensor parallel), no P2P · PP2 (layer partition 25,23), no P2P · PP2 with PLE CPU offload · PP2, MTP 3, no P2P · PP2, MTP depth 3 · PP3 · PP3, no P2P (works at PP2 on two cards) · PP8 (also PP9 and PP10), TP1 · PP8, partition 5,5,5,5,5,5,5,5 · PP8, TP1 · TP2 over NCCL (GPU P2P), DFlash2 K7 plus hybrid drafting · TP2 with expert parallel · TP2 with whole-expert expert parallel · TP2 x PP3, partition 14,14,12 · TP4 · TP4 (default) or PP4 · TP4 (default) or PP4 for x1/x4 links · TP4 (P2P on); expert parallel off · TP4 with expert parallel (96 whole experts per card) · TP4 with expert parallel, MTP 3 · TP4 x PP2, partition 20,20 · TP4, P2P on · TP4, peer access off (NCCL_P2P_DISABLE=1), host all-reduce · TP5 across four CMP 170HX and one RTX PRO 6000 Blackwell (native backend, reduces through host memory); MTP draft on the PRO · TP6 (virtual-head padding), PP1 · Other · None

262,144 (2) · 262,144 advertised (tested through 32K) (1) · 262,144 native (YaRN to about 500k in the recipe's long variant) (1)

110.99 (1) · 133.39 (1) · 170 (1) · 296.4 (1)

None (1) · 2,636 (1) · 4,243 (1) · 6,213.66 (1)

64 GB (1) · 64 GiB, --gpu-memory-utilization 0.90 (1) · 64 GiB, --gpu-memory-utilization 0.95 (fixed) (1) · 64 GiB, --gpu-memory-utilization 0.96 (1)

None (1) · 140 W (200 W observed-draw abort guard) (1) · 180 W (2)

None (1) · 580 or newer (measured on 610.57.04) (1) · 610.43.03 (1) · NVIDIA 610.43.02 (1)

None (1) · amoghmunikote/cmpunlocker 6c442eeb (patch 0016 SM reconfig, 74 SMs) plus Bayley P2P patches 0011/0013/0015 (1) · cmpunlocker (64 GiB) (1) · cmpunlocker (64 GiB, 74 SMs) (1)

Proven (1) · Reported (3)

dufrelic (1) · iIIusi0n (1) · kellertreppe123 (1) · nathang6866 (1)

September 19, 2026 (1) · October 3, 2026 (1) · October 5, 2026 (1) · October 23, 2026 (1)

https://github.com/DUFRelic/qwen38-flash-next-recipe (1) · https://github.com/ElectroHobo/qwen38-flash-next-3bpw-1x-cmp170hx-recipe (1) · https://github.com/iIIusi0n/qwen3.8-flash-next-cmp170hx (1) · https://www.reddit.com/r/LocalLLM/comments/1w2f03e/qwen3827b_w4a16_on_one_64_gb_cmp_170hx_16070/ (1)