Drilldown: Recipe

DeepSeek V4.1 Flash on 4x 8 GB plus RTX PRO 6000 (ExLlamaV3 TP5) (1) · Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2) (1)

DeepSeek-V4.1-Flash-1017 (1) · Qwen3.8-Flash-Next (125B MoE per the recipe) (1)

EXL3 3.5 bpw (LibertAIDAI; routed experts 4 bpw on layers 0-9 and 30-39, 3 bpw on 10-29; attention and shared experts 6 bpw) plus the official Engram shards (1) · W4A16: halt95/Qwen3.8-Flash-Next-W4A16-Merlin (AutoRound INT4 g128 experts, INT8 GDN projections, FP8 n-gram table) (1)

Morrowmake/vllm-cmp170hx (vLLM fork, container image) · vLLM PR #56201 head (79a7108d) plus the two Deepinfer-SM80 patches and 30 drop-in modules · Custom SM80 vLLM (patched from public commit 79a7108d9) · ExLlamaV3 (patched, pinned) · ExLlamaV3 1.5.4 (built for SM80) with a small OpenAI-compatible server · ExLlamaV3 fork (anoane/exllamav3-anemone) + TabbyAPI fork (anoane/tabbyAPI) · GNmac-lab recipe image 1.2.0 (vLLM with sm_80 patches) plus three patches · Morrowmake/vllm-cmp170hx (bind-mounted sm80 package) plus the author's tensor-core MoE kernels · Morrowmake/vllm-cmp170hx (EXL3 path) · Morrowmake/vllm-cmp170hx (via the glm53-flash-cmp170hx-recipe) · Morrowmake/vllm-cmp170hx branch ampere-glm53 plus fail-closed source patches · Morrowmake/vllm-cmp170hx image (Qwen EXL3 release) with patch files mounted read-only · Morrowmake/vllm-cmp170hx image (tag qwen38-recipe-1.0.0) plus one mounted file (vLLM #58235 vision-sink fix) · Morrowmake/vllm-cmp170hx image with patch files mounted read-only · Project Maya (Strata and llama.cpp lineage) · Project Maya (Strata and llama.cpp lineage, EXL3 loading, DFlash2 drafter) · vLLM · vLLM (build not stated) · vLLM 0.29.1rc1 nightly plus the checkpoint author's patches · vLLM a5a30471 plus a 13-file downstream patch (native Linux AIO PLE reader) · vLLM nightly + PR #53899 (a5530b9) + the repo's patch scripts · vLLM nightly 7d0b4e57 plus the author's pure-Python fork clee/vllm-sm80 (branch q38fn/stack) · vLLM PR #56201 head (79a7108d) plus a 20-patch sm_80 series; image zanooda/vllm-sm80-ds41f:v41-sm80 · vLLM v0.30.0 + zankich INT4-PLE plugin + six hash-guarded patch layers · vLLM with a pinned sm_80 patch stack (built locally from source) · wtdcode/vllm-backport (image lazymio/vllm-backport:latest-sm80) · wtdcode/vllm-backport 3142ded1 plus a two-file patch and two tinybeast PLE plugins (venv, no container) · Other · None

recipe 0.3.1; engine 18305948ac (1) · tag recipe-09 (309ef6d1); TabbyAPI deepseek-v41 (ce250ff) (1)

8 GB · Other · None

2 (1) · 4 (1)

TP2 with whole-expert expert parallel (1) · TP5 across four CMP 170HX and one RTX PRO 6000 Blackwell (native backend, reduces through host memory); MTP draft on the PRO (1)

1,048,576 FP16 (1) · 262,144 (1)

99.9 (1) · 349.7 (1)

2,116 (1) · 4,696 (1)

64 GiB, GPU_UTIL 0.95 (fixed) (1) · gpu_split 58 GB per CMP, 70 GB on the PRO (1)

150 W per CMP (PRO 400 W) (1) · 180 W (1)

580 or newer (measured on 610.57.04); open modules with HMM for the table-on-disk mode (1) · NVIDIA open 610.57.04 patched with Morrowmake/cmpunlocker 1f2c498 (profile 8gb) (1)

cmpunlocker (64 GiB, 74 SMs) with a P2P-capable build (1) · Morrowmake/cmpunlocker 1f2c498; modprobe RmForceEnableGen2=1, RMPcieLinkSpeed=0x1, ForceP2P=0x11 (1)

Proven (1) · Reported (1)

anoane (1) · nathang6866 (1)

October 20, 2026 (1) · October 23, 2026 (1)

https://github.com/anoane/09-deepseek-v41-flash-1017-exl3-tp5-1m (1) · https://github.com/ElectroHobo/qwen38-flash-next-w4a16-2x-cmp170hx-recipe (1)