Drilldown: Recipe

DeepSeek V4.1 Flash on 4x 8 GB plus RTX PRO 6000 (ExLlamaV3 TP5) (1) · MiMo V2.6 Flash on 4x 8 GB (1) · Qwen3.8 Flash Next on 1x 8 GB (ElectroHobo 3 bpw) (1) · Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2) (1)

DeepSeek-V4.1-Flash-1017 (1) · MiMo-V2.6-Flash-RL (309B MoE, 15B active; text, image, video and audio input) (1) · Qwen3.8-Flash-Next (125B MoE per the recipe) (2)

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) · 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) · NVFP4 (ModelOpt mixed precision): ProCreations/MiMo-V2.6-Flash-RL-NVFP4, run as 4-bit weights with BF16 activations through Marlin (1) · W4A16: halt95/Qwen3.8-Flash-Next-W4A16-Merlin (AutoRound INT4 g128 experts, INT8 GDN projections, FP8 n-gram table) (1)

ExLlamaV3 fork (anoane/exllamav3-anemone) + TabbyAPI fork (anoane/tabbyAPI) (1) · Morrowmake/vllm-cmp170hx image (Qwen EXL3 release) with patch files mounted read-only (1) · Morrowmake/vllm-cmp170hx image (tag qwen38-recipe-1.0.0) plus one mounted file (vLLM #58235 vision-sink fix) (1) · Morrowmake/vllm-cmp170hx image with patch files mounted read-only (1)

recipe 0.2.1; engine 18305948ac (2) · recipe 0.3.1; engine 18305948ac (1) · tag recipe-09 (309ef6d1); TabbyAPI deepseek-v41 (ce250ff) (1)

8 GB (4)

1 (1) · 2 (1) · 4 (2)

single card (1) · TP2 with whole-expert expert parallel (1) · TP4 (P2P on); expert parallel off (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 (1) · 1,048,576 FP16 (1) · 262,144 (2)

99.9 (1) · 274.9 (1) · 296.4 (1) · 349.7 (1)

2,116 (1) · 2,722 (1) · 4,243 (1) · 4,696 (1)

64 GiB, --gpu-memory-utilization 0.95 (fixed) (1) · 64 GiB, GPU_UTIL 0.95 (fixed) (1) · 64 GiB; weights 45.5 GiB plus 10 GiB KV per card; GPU_UTIL 0.94 (fixed) (1) · gpu_split 58 GB per CMP, 70 GB on the PRO (1)

150 W per CMP (PRO 400 W) (1) · 180 W (3)

580 or newer (measured on 610.57.04) (2) · 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) (1) · cmpunlocker (64 GiB, 74 SMs) with a P2P-capable build (2) · Morrowmake/cmpunlocker 1f2c498; modprobe RmForceEnableGen2=1, RMPcieLinkSpeed=0x1, ForceP2P=0x11 (1)

Proven (1) · Reported (3)

nathang6866 · voidgiraffe · HalfVulpes · 1a2b3c0967 · 68x0 · alanoo · anoane · c133 · CurtisAccelerate · darkmaniac7 · dufrelic · greencandle9753 · iIIusi0n · kellertreppe123 · matstatman · mrmannyatx · nestaa51 · pcsv. · plean. · reala7096 · shirogeek · vektor_prime · wgkbfcp · ZackO2o · Zanooda · Other · None

October 18, 2026 (1) · October 20, 2026 (1) · October 23, 2026 (2)

https://github.com/anoane/09-deepseek-v41-flash-1017-exl3-tp5-1m (1) · https://github.com/ElectroHobo/mimo-v26-flash-cmp170hx-recipe (1) · https://github.com/ElectroHobo/qwen38-flash-next-3bpw-1x-cmp170hx-recipe (1) · https://github.com/ElectroHobo/qwen38-flash-next-w4a16-2x-cmp170hx-recipe (1)