Drilldown: Recipe

DeepSeek V4 Flash 0731 on 3x 8 GB (PP3 DSpark) (1) · DeepSeek V4.1 Flash on 4x 8 GB (alanoo fidelity build) (1) · DeepSeek V4.1 Flash on 4x 8 GB plus RTX PRO 6000 (ExLlamaV3 TP5) (1) · GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) (1) · GLM 5.3 Flash on 4x 8 GB (kaka86mm EXL3 + Marlin sidecar + DFlash2) (1) · Qwen3.8 Flash Next on 3x or 4x 8 GB (zebgop-ops FP8 PP) (1)

DeepSeek-V4-Flash-0731 (1) · DeepSeek-V4.1-Flash-1017 (2) · GLM-5.3-Flash (320B MoE, 18B active) (1) · GLM-5.3-Flash (320B MoE, 18B active, multimodal) (1) · Qwen/Qwen3.8-Flash-Next-FP8 (revision bcd9f01d) (1)

Native MXFP4 experts + FP8 attention (official deepseek-ai checkpoint, about 155.4 GiB); KV cache fp8 · Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 · AutoRound 3 bpw + MTP: klee100/Qwen3.8-Flash-Next-AutoRound-3bpw-MTP; 95.37 GiB BF16 PLE table on SSD · AutoRound mixed 2/3-bit (klee100, 2.998 effective bpw, 47 GiB on card) plus halt95's FP8 n-gram table in host RAM · AWQ INT4: cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4; 95 GiB BF16 PLE table in host RAM (FP8 table optional) · AWQ W4A16 g32 (leoncca/Qwen3.8-Flash-Next-Uncensored-AWQ-g32); PLE table converted to INT4 and served from SSD · AWQ W4A16: wtdcode/Qwen3.8-Flash-Next-AWQ-W4A16 · EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3 · EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3; DFlash2 drafter converted to EXL3 6 bpw · 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 · EXL3 4 bpw (Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw, 164 GB) plus an offline-generated Marlin INT4 decode sidecar (151 GB) · EXL3 5.05 bpw; PLE (n-gram) table read from SSD · GLM-5.3-Int4-Int8Mix (INT4 / INT8 mix); KV fp8_ds_mla · GLM-5.3-Int4-Int8Mix-AWQ-g64; drafter incoai/GLM-5.3-DFlash2 · Hybrid: all 40 expert layers at 3 bits (LibertAI 2.9 bpw pack with layers 18-22 grafted from the 3.5 bpw pack), official FP8 dense tensors · Hybrid: LibertAI EXL3 2.9 bpw routed experts (layers 0-39), official FP8/FP4 for attention, shared experts, head, draft and Engram; KV fp8_ds_mla · Hybrid: routed experts EXL3 (mostly 3 bpw, 2 bpw on layers 18-22) from LibertAI's 2.9 bpw pack; attention, shared experts, indexer and DSpark drafter official FP8; KV fp8_ds_mla · Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16 with the 38 heretic attn.wo_b tensors · Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16; KV BF16 · LibertAIDAI/GLM-5.3-Flash-NVFP4 (ModelOpt NVFP4, routed experts only, 181 GiB), run through Marlin W4A16 kernels; GLM-5.3-Flash DFlash2 drafter (5 sliding-window layers) · Maya GGUF (Maya-S 340 GB or Maya-XS 277 GB; Engram tables stay on SSD) · Maya-M GGUF (116 GB), every expert in VRAM; Maya-S (96.5 GB) and Maya-S24 (94.7 GB) also reported · Mixed INT4 AutoRound (Minachist/Qwen3.8-Flash-Next-INT4-Mixed-AutoRound, about 65 GB) with the full BF16 PLE table on NVMe · NVFP4 (ModelOpt mixed precision): ProCreations/MiMo-V2.6-Flash-RL-NVFP4, run as 4-bit weights with BF16 activations through Marlin · NVFP4 main weights through Marlin (no native FP4 on sm_80), mixed FP8 and BF16 MTP weights, FP8 PLE table (about 51.2 GB) kept on NVMe · Official block FP8, about 126 GiB including the about 50 GiB FP8 n-gram (PLE) table, which is served from host RAM; KV cache bf16 (the QSA layers refuse fp8) · Official checkpoint (510 GB: MXFP4 routed experts, FP8 dense layers, 203 GB FP8 Engram table); KV cache fp8_ds_mla · Official checkpoint as downloaded: MXFP4 routed experts as Marlin W4A16, MXFP8 dense as Marlin W8A16; KV fp8 · Official checkpoint: dense MXFP8 through an SM80 Marlin W8A16 path, routed experts MXFP4 through W4A16; KV FP8 · W4A16 (INT4 weights, group size 128, FP16 activations): canada-quant/GLM-5.3-Flash-W4A16-MTP · W4A16 (INT4 weights, Marlin kernels) · W4A16 AutoRound: Intel/Qwen3.8-Flash-Next-W4A16-AutoRound; 95.4 GiB BF16 PLE table streamed from NVMe; KV fp8 e4m3 (software-decoded on SM80) · W4A16 INT4 g128 with the PLE table repacked INT4 g32: zankich/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV; KV BF16 · W4A16, about 4.3 bpw, 163 GiB on disk · W4A16: halt95/Qwen3.8-Flash-Next-W4A16-Merlin (AutoRound INT4 g128 experts, INT8 GDN projections, FP8 n-gram table) · Other · None

ExLlamaV3 (patched, pinned) (1) · ExLlamaV3 fork (anoane/exllamav3-anemone) + TabbyAPI fork (anoane/tabbyAPI) (1) · GNmac-lab recipe image 1.2.0 (vLLM with sm_80 patches) plus three patches (1) · vLLM: haosdent/vllm branch dsv4-flash-a100 with a vendored copy of allover326's 7 patches (image dsv4-a100:devel) (1) · vLLM: hyd998877/vllm_170hx_glm53-flash-exl3-optimized with kaka86mm's overlay patches; exllamav3 1.4.8 (1) · vLLM: the vllm/vllm-openai:qwen38-flash-next image with zebgop-ops patches 0001 to 0016 (never 0008-WITHDRAWN) and three ported files (1)

0.1.dev20073+g8e685d198; repo commit 2f551bb (patch 0016, 2026-10-02) (1) · f8ea5bb-era base with the allover326 patch set as of mid-August 2026 (1) · final production build v30 (numbers measured on v24); repo commit 218b47f (1) · GNmac-lab at 2ad5741ab1; image ds41-cmp170hx:1.2.0-fidelity (1) · tag recipe-09 (309ef6d1); TabbyAPI deepseek-v41 (ce250ff) (1) · upstream dev 0f685aa9fd plus 9 patches; repo db01c256 (1)

8 GB (6)

2 (1) · 3 (1) · 4 (4)

PP3 (VLLM_PP_LAYER_PARTITION=15,15,13) with DSpark, 5 draft tokens; PCIe Gen2 x4, no P2P (1) · PP4 (VLLM_PP_LAYER_PARTITION=13,11,11,10), DFlash2 k=7; PCIe Gen2 x4, no P2P (1) · PP4, VLLM_PP_LAYER_PARTITION=12,12,12,12, MTP 3, FULL_AND_PIECEWISE CUDA graphs; PCIe Gen2 x4 per card, no P2P. Three cards: PP3 with 16,17,15 or 16,16,16 (1) · TP2 over NCCL (GPU P2P), DFlash2 K7 plus hybrid drafting (1) · TP4, P2P on (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) · 262,144 native; 524,288 with YaRN 2.0; 1,048,576 with YaRN 4.0 (needle verified on PP3) (1) · 392,960 per request; 655,360-token Q8 KV pool shared by two requests (1) · 524,288 (1) · 691,968 (largest stable context at --gpu-memory-utilization 0.965) (1)

80 - 110 (6)

None (1) · 1,533 (1) · 2,077 (1) · 2,116 (1) · 2,514 (1) · 11,100 (1)

64 GB (65,536 MiB), --gpu-memory-utilization 0.965 (1) · 64 GB, --gpu-memory-utilization 0.93 (0.94 is the ceiling) (1) · 64 GiB (64,911 MiB usable), GPU_MEMORY_UTILIZATION 0.98 (1) · 64 GiB; --gpu-memory-utilization 0.94 plus per-rank KV budgets; measured 34.0-34.3 GiB per 12-layer rank, 37.9 GiB on the drafter rank (1) · about 62.9 / 62.3 GiB idle, peaks about 64.4 / 63.0 GiB with vision (1) · gpu_split 58 GB per CMP, 70 GB on the PRO (1)

100 W (1) · 150 W per CMP (PRO 400 W) (1) · 180 W (author's cards); about 170 W with a modest memory OC in his Discord numbers (1) · 200 W (3)

None (1) · 610.43.02 (nvidia-driver-610-open) (1) · 610.43.03 (open modules) (1) · 610.57.04 (1) · NVIDIA open 610.43.02 from a cmpunlocker fork with a P2P patch (1) · NVIDIA open 610.57.04 patched with Morrowmake/cmpunlocker 1f2c498 (profile 8gb) (1)

cmpunlocker (64 GB) (1) · cmpunlocker (64 GiB); P2P-capable driver for the NCCL backend (1) · cmpunlocker 360acd7 (2026-08-06), 8 GB profile (1) · cmpunlocker fork with P2P (without Morrowmake's TRAP31 patch) (1) · cmpunlocker VRAM unlock (64 GiB per card) (1) · Morrowmake/cmpunlocker 1f2c498; modprobe RmForceEnableGen2=1, RMPcieLinkSpeed=0x1, ForceP2P=0x11 (1)

Reported · Proven · Other · None

alanoo (1) · anoane (1) · dkpoulsen (1) · flipnburner (1) · nestaa51 (1) · tim-odonnell (1)

August 2026 (1) · September 2026 (2) · October 2026 (3)

https://github.com/alanoo81/deepseek-v41-flash-4x-cmp170hx (1) · https://github.com/alexvnesta/glm53-flash-cmp170hx-tp (1) · https://github.com/anoane/09-deepseek-v41-flash-1017-exl3-tp5-1m (1) · https://github.com/kaka86mm/glm53-flash-170hx (1) · https://github.com/tim-odonnell/cmp170hx-deepseek-v4-flash (1) · https://github.com/zebgop-ops/qwen38-flashnext-pp (1)