Drilldown: Recipe

DeepSeek V4 Flash 0731 on 3x 8 GB (PP3 DSpark) (1) · DeepSeek V4 Flash 0731 on 4x 8 GB (allover326 PP4 DSpark) (1) · DeepSeek V4.1 Flash on 2x 8 GB (Project Maya) (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) · DeepSeek V4.1 Flash on 6x 8 GB (TP6) (1) · DeepSeek V4.1 Flash on 8x 8 GB (Zanooda PP8) (1) · GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 layer split) (1) · GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) (1) · GLM 5.3 Flash on 2x 8 GB (Project Maya) (1) · GLM 5.3 Flash on 4x 8 GB (kaka86mm EXL3 + Marlin sidecar + DFlash2) (1) · GLM 5.3 Flash on 8x 8 GB (promisezackr PP8 NVFP4 DFlash2) (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 Flash Next on 2x 8 GB (vektorprime PP2) (1) · Qwen3.8 Flash Next on 3x 8 GB (clee PP3) (1) · Qwen3.8 Flash Next on 3x or 4x 8 GB (zebgop-ops FP8 PP) (1) · Qwen3.8 Flash Next on 4x 8 GB (1)

DeepSeek-V4-Flash-0731 (2) · DeepSeek-V4.1-Flash-1017 (4) · DeepSeek-V4.1-Flash-1017 (748B in the Maya README) (1) · GLM-5.3-Flash (1) · GLM-5.3-Flash (320B MoE, 18B active) (2) · GLM-5.3-Flash (320B MoE, 18B active, multimodal) (1) · GLM-5.3-Flash (321B in the Maya README) (1) · Qwen/Qwen3.8-Flash-Next-FP8 (revision bcd9f01d) (1) · Qwen3.8-Flash-Next (5)

AutoRound 3 bpw + MTP: klee100/Qwen3.8-Flash-Next-AutoRound-3bpw-MTP; 95.37 GiB BF16 PLE table on SSD (1) · AWQ INT4: cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4; 95 GiB BF16 PLE table in host RAM (FP8 table optional) (1) · AWQ W4A16: wtdcode/Qwen3.8-Flash-Next-AWQ-W4A16 (1) · EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3 (1) · EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3; DFlash2 drafter converted to EXL3 6 bpw (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) · EXL3 4 bpw (Mia-AiLab/GLM-5.3-Flash-EXL3-TR3-4bpw, 164 GB) plus an offline-generated Marlin INT4 decode sidecar (151 GB) (1) · 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 (1) · 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) (1) · Maya GGUF (Maya-S 340 GB or Maya-XS 277 GB; Engram tables stay on SSD) (1) · Maya-M GGUF (116 GB), every expert in VRAM; Maya-S (96.5 GB) and Maya-S24 (94.7 GB) also reported (1) · Native MXFP4 experts + FP8 attention (official deepseek-ai checkpoint, about 155.4 GiB); KV cache fp8 (2) · 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) (1) · Official checkpoint as downloaded: MXFP4 routed experts as Marlin W4A16, MXFP8 dense as Marlin W8A16; KV fp8 (1) · Official checkpoint: dense MXFP8 through an SM80 Marlin W8A16 path, routed experts MXFP4 through W4A16; KV FP8 (1) · 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) (1) · W4A16 INT4 g128 with the PLE table repacked INT4 g32: zankich/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV; KV BF16 (1)

Custom SM80 vLLM (patched from public commit 79a7108d9) (1) · ExLlamaV3 (patched, pinned) (1) · ExLlamaV3 1.5.4 (built for SM80) with a small OpenAI-compatible server (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) · Project Maya (Strata and llama.cpp lineage) (1) · Project Maya (Strata and llama.cpp lineage, EXL3 loading, DFlash2 drafter) (1) · vLLM a5a30471 plus a 13-file downstream patch (native Linux AIO PLE reader) (1) · vLLM nightly + PR #53899 (a5530b9) + the repo's patch scripts (1) · vLLM nightly 7d0b4e57 plus the author's pure-Python fork clee/vllm-sm80 (branch q38fn/stack) (1) · vLLM PR #56201 head (79a7108d) plus a 20-patch sm_80 series; image zanooda/vllm-sm80-ds41f:v41-sm80 (1) · vLLM v0.30.0 + zankich INT4-PLE plugin + six hash-guarded patch layers (1) · vLLM: haosdent/vllm branch dsv4-flash-a100 plus allover326's 7 patches (DSpark under PP, sm_80 prefill top-k fallback, logits row chunking) (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: official vllm/vllm-openai:glm53-flash image plus promisezackr's 24 patches (SM8x base from 344303947/vllm170hx, pipeline parallel for GLM-5.3-Flash, DFlash2 under PP8) (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) · wtdcode/vllm-backport (image lazymio/vllm-backport:latest-sm80) (1)

None (1) · 0.1.dev1+ga5530b90c.d20260829 (1) · 0.1.dev20073+g8e685d198; repo commit 2f551bb (patch 0016, 2026-10-02) (1) · 0.29.1rc1.dev402+ga5a30471f.ple1 (1) · base c3046d1 (recommended since 2026-08-13; first released on f8ea5bb); repo commit 3dd2d88 (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) · Flun/glm53-flash-cmp170hx-exl3 at 5933112 (1) · GNmac-lab at 2ad5741ab1; image ds41-cmp170hx:1.2.0-fidelity (1) · latest-sm80 tag as of 2026-10-16 (1) · q38fn/stack on nightly-7d0b4e57 (1) · recipe 31318fd; patches 0001-0020 (1) · reports 0.0.0; patches/public-base-to-runtime.patch (109 files) (1) · tag recipe-09 (309ef6d1); TabbyAPI deepseek-v41 (ce250ff) (1) · upstream dev 0f685aa9fd plus 9 patches; repo db01c256 (1) · v0.1.dev20051+g487ecf187 (the image's vLLM); repo commit 90ec72e (1) · v1.0.6 (this row) (1) · vLLM v0.30.0 (ced6857a); plugin zankich/vllm v0.29.0z (6609bfef) (1)

8 GB (18)

1 - 2 (1) · 2 - 3 (6) · 3 - 4 (2) · 4 - 8 (9)

engine-managed expert tiering across VRAM, host RAM and NVMe (1) · engine-managed split with expert tiering across VRAM, RAM and NVMe (1) · layer split (no tensor parallel), no P2P (1) · PP2 with PLE CPU offload (1) · PP2, MTP depth 3 (1) · PP3 (VLLM_PP_LAYER_PARTITION=15,15,13) with DSpark, 5 draft tokens; PCIe Gen2 x4, no P2P (1) · PP3, no P2P (works at PP2 on two cards) (1) · PP4 (VLLM_PP_LAYER_PARTITION=12,12,12,7) 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) · PP8 (VLLM_PP_LAYER_PARTITION=7,5,7,5,7,5,7,2) with DFlash2, 7 draft tokens; PCIe Gen2 x4, no P2P; host 2x Xeon E5-2690 v4, 96 GB DDR4 (1) · PP8, partition 5,5,5,5,5,5,5,5 (1) · single card (1) · TP2 over NCCL (GPU P2P), DFlash2 K7 plus hybrid drafting (1) · TP4 with expert parallel, MTP 3 (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) · TP6 (virtual-head padding), PP1 (1)

None (2) · 1,000,000 (YaRN x4 from 262,144) (1) · 1,047,736 verified one-shot; 1,002,852 over a 405-turn chat with DSV4_LOGITS_ROW_CHUNK=64 (1) · 1,048,576 (2) · 1,048,576 (KV pool 6.62M tokens at utilization 0.90, about 6.3 concurrent 1M requests) (1) · 1,048,576 (YaRN x4) (1) · 1,048,576 FP16 (1) · 262,144 (3) · 262,144 advertised (tested through 32K) (1) · 262,144 native; 524,288 with YaRN 2.0; 1,048,576 with YaRN 4.0 (needle verified on PP3) (1) · 392,960 (profile q8_384k) (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)

60 - 70 (3) · 70 - 100 (8) · 100 - 108 (2) · 108 - 140 (5)

None (6) · 0 - 2,000 (2) · 2,000 - 5,000 (6) · 5,000 - 12,000 (4)

64 GB (3) · 64 GB (65,536 MiB), --gpu-memory-utilization 0.965 (1) · 64 GB, --gpu-memory-utilization 0.85 (1) · 64 GB, --gpu-memory-utilization 0.90 (1) · 64 GB, --gpu-memory-utilization 0.93 (0.94 is the ceiling) (1) · 64 GB, --gpu-memory-utilization 0.95 (1) · 64 GB; weights 35.9 GB per rank (ranks 1-6), 37.6 GB rank 0, 46.3 GB rank 7 (1) · 64 GiB (64,911 MiB usable), GPU_MEMORY_UTILIZATION 0.98 (1) · 64 GiB, --gpu-memory-utilization 0.90 (1) · 64 GiB, --gpu-memory-utilization 0.96 (1) · 64 GiB, gpu-memory-utilization 0.92 (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) · 64 GiB; weights PP0 39.48 GiB, PP1 45.53 GiB; --gpu-memory-utilization 0.9 (1) · about 58 GiB of 64 GiB (116.6 GiB resident across both) (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)

None (8) · 100 W (1) · 140 W (200 W observed-draw abort guard) (1) · 150 W per CMP (PRO 400 W) (1) · 180 W (2) · 180 W (author's cards); about 170 W with a modest memory OC in his Discord numbers (1) · 180 W (one card unstable at 250 W) (1) · 200 W (3)

None (11) · 610.43.02 (nvidia-driver-610-open) (1) · 610.43.03 (1) · 610.43.03 (open modules) (1) · 610.57.04 (1) · NVIDIA 610.43.02 (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)

None (7) · amoghmunikote/cmpunlocker 6c442eeb (patch 0016 SM reconfig, 74 SMs) plus Bayley P2P patches 0011/0013/0015 (1) · cmpunlocker (64 GB) (1) · cmpunlocker (64 GiB) (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 GB per card) (1) · cmpunlocker VRAM unlock (64 GiB per card) (1) · cmpunlocker VRAM unlock (65,536 MiB per card) (1) · cmpunlocker with the BAR1 P2P patches (64 GiB BAR1 per card) (1) · Morrowmake/cmpunlocker 1f2c498; modprobe RmForceEnableGen2=1, RMPcieLinkSpeed=0x1, ForceP2P=0x11 (1)

Proven (4) · Reported (14)

68x0 (1) · alanoo (1) · anoane (1) · c133 (1) · deejay9983 (1) · dkpoulsen (1) · dufrelic (1) · flipnburner (1) · iIIusi0n (1) · matstatman (1) · nestaa51 (1) · pcsv. (1) · promisezackr (1) · tim-odonnell (1) · vektor_prime (1) · wgkbfcp (1) · ZackO2o (1) · Zanooda (1)

August 2026 (4) · September 2026 (3) · October 2026 (10)

https://discord.com/channels/1518634379604136178/1528826330454429706/1560602001908170753 (1) · https://github.com/alanoo81/deepseek-v41-flash-4x-cmp170hx (1) · https://github.com/alexvnesta/glm53-flash-cmp170hx-tp (1) · https://github.com/allover326/deepseek-v4-cmp170hx (1) · https://github.com/anoane/09-deepseek-v41-flash-1017-exl3-tp5-1m (1) · https://github.com/clee/qwen3.8-flash-next-3x170hx (1) · https://github.com/DUFRelic/qwen38-flash-next-recipe (1) · https://github.com/Flun/glm53-flash-cmp170hx-exl3 (1) · https://github.com/iIIusi0n/qwen3.8-flash-next-cmp170hx (1) · https://github.com/kaka86mm/glm53-flash-170hx (1) · https://github.com/mw00/project-maya#how-fast-is-it (2) · https://github.com/promisezackr/glm53-flash-170hx-pp8 (1) · https://github.com/tim-odonnell/cmp170hx-deepseek-v4-flash (1) · https://github.com/vektorprime/qwen38-flash-next-pp2 (1) · https://github.com/wgkbfcp/deepseek-v4.1-flash-1017-cmp170hx-recipe (1) · https://github.com/Zanooda/deepseek-v4.1-flash-cmp170hx (1) · https://github.com/zebgop-ops/qwen38-flashnext-pp (1)