Drilldown: Recipe

DeepSeek V4.1 Flash heretic on 4x 8 GB (1) · DeepSeek V4.1 Flash on 6x 8 GB (Deepinfer-SM80 TP2 x PP3) (1) · DeepSeek V4.1 Flash on 6x 8 GB (TP6) (1) · DeepSeek V4.1 Flash on 8x 8 GB (Deepinfer-SM80 TP4 x PP2) (1) · GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 layer split) (1) · GLM 5.3 Flash on 2x 8 GB (Project Maya) (1) · GLM 5.3 Flash on 3x 8 GB (1) · Qwen3.8 Flash Next on 2x 8 GB (Morrowmake EXL3) (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 4x 8 GB (1)

DeepSeek-V4.1-Flash-1017 (3) · DeepSeek-V4.1-Flash-1017 heretic (pqhaz abliteration) (1) · GLM-5.3-Flash (320B MoE, 18B active) (2) · GLM-5.3-Flash (321B in the Maya README) (1) · Qwen3.8-Flash-Next (4)

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 5.05 bpw; PLE (n-gram) table read from SSD (1) · Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16 with the 38 heretic attn.wo_b tensors (1) · Maya-M GGUF (116 GB), every expert in VRAM; Maya-S (96.5 GB) and Maya-S24 (94.7 GB) also reported (1) · Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 (2) · 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, about 4.3 bpw, 163 GiB on disk (1)

Custom SM80 vLLM (patched from public commit 79a7108d9) (1) · ExLlamaV3 1.5.4 (built for SM80) with a small OpenAI-compatible server (1) · Morrowmake/vllm-cmp170hx (EXL3 path) (1) · Morrowmake/vllm-cmp170hx (via the glm53-flash-cmp170hx-recipe) (1) · Morrowmake/vllm-cmp170hx (vLLM fork, container image) (1) · Project Maya (Strata and llama.cpp lineage, EXL3 loading, DFlash2 drafter) (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 the two Deepinfer-SM80 patches and 30 drop-in modules (2) · wtdcode/vllm-backport (image lazymio/vllm-backport:latest-sm80) (1)

None · recipe 0.2.1; engine 18305948ac · 0.1.dev1+ga5530b90c.d20260829 · 0.29.1rc1.dev402+ga5a30471f.ple1 · Deepinfer-SM80 2a828fc lineage; vLLM v0.28.1rc0 base · Deepinfer-SM80, vLLM v0.28.1rc0 base · Flun/glm53-flash-cmp170hx-exl3 at 5933112 · GNmac-lab at 2ad5741ab1; image ds41-cmp170hx:1.2.0-fidelity · latest-sm80 tag as of 2026-10-16 · not stated · q38fn/stack on nightly-7d0b4e57 · recipe 0.3.1; engine 18305948ac · recipe 1.0.0 (commit d4d1f525), engine a0c57b52 · recipe 1.0.0, engine a0c57b5237 · recipe 1.2.0 · recipe 1.8.0, engine 819146023f · recipe 31318fd; patches 0001-0020 · reports 0.0.0; patches/public-base-to-runtime.patch (109 files) · tag qwen38-recipe-1.0.0, engine 18305948ac · tag recipe-09 (309ef6d1); TabbyAPI deepseek-v41 (ce250ff) · upstream dev 0f685aa9fd plus 9 patches; repo db01c256 · v1.0.6 (this row) · vLLM 0.30.1rc1, fork commit 378c37b0 · vLLM v0.30.0 (ced6857a); plugin zankich/vllm v0.29.0z (6609bfef) · vllm-backport 3142ded1 · vllm-cmp170hx 18305948ac; TC-MoE fb9786ec · Other

8 GB (11)

2 (4) · 3 (2) · 4 (2) · 6 (2) · 8 (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) · PP3 (1) · PP3, no P2P (works at PP2 on two cards) (1) · TP2 with expert parallel (1) · TP2 x PP3, partition 14,14,12 (1) · TP4 (1) · TP4 with expert parallel, MTP 3 (1) · TP4 x PP2, partition 20,20 (1) · TP6 (virtual-head padding), PP1 (1)

None (1) · 1,000,000 (YaRN x4 from 262,144) (1) · 1,048,576 (3) · 1,048,576 (YaRN x4) (1) · 262,144 (3) · 392,960 (profile q8_384k) (1) · about 1,000,000 (8-bit KV cache) (1)

None (1) · 60 - 80 (5) · 80 - 170 (5)

None (4) · 0 - 2,000 (1) · 2,000 - 5,000 (6)

64 GB (2) · 64 GB, --gpu-memory-utilization 0.95 (1) · 64 GiB (2) · 64 GiB, --gpu-memory-utilization 0.92 (miner co-resident) (1) · 64 GiB, --gpu-memory-utilization 0.95 (2) · 64 GiB, gpu-memory-utilization 0.92 (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)

None · 180 W · 200 W · 140 W · 140 W (200 W observed-draw abort guard) · 150 W per CMP (PRO 400 W) · 170 W (author's 10-card run) · 180 W (author's cards); about 170 W with a modest memory OC in his Discord numbers · 180 W (nvidia-smi -pl 180) · 180 W (one card unstable at 250 W) · Other

None · 580 or newer (measured on 610.57.04) · NVIDIA 580 or newer; P2P needs the Morrowmake/cmpunlocker build · 580 or newer (measured on 610.57.04); open modules with HMM for the table-on-disk mode · 610.43.03 · 610.57.04 · CUDA 13 capable; measured on 610.43.02 · NVIDIA 610.43.02 · NVIDIA open 610.43.02 from a cmpunlocker fork with a P2P patch · NVIDIA open 610.57.04 patched with Morrowmake/cmpunlocker 1f2c498 (profile 8gb) · nvidia-open 610.43.02 + cmpunlocker v0.5 or newer · Other

None (11)

Proven (2) · Reported (9)

68x0 (1) · c133 (1) · darkmaniac7 (1) · HalfVulpes (2) · matstatman (1) · shirogeek (1) · vektor_prime (1) · voidgiraffe (1) · wgkbfcp (1) · ZackO2o (1)

August 2026 (1) · October 2026 (9)

https://discord.com/channels/1518634379604136178/1518695289278566450/1561743826945843304 (1) · https://discord.com/channels/1518634379604136178/1528826330454429706/1560602001908170753 (1) · https://github.com/clee/qwen3.8-flash-next-3x170hx (1) · https://github.com/Flun/glm53-flash-cmp170hx-exl3 (1) · https://github.com/HalfVulpes/Deepinfer-SM80 (2) · https://github.com/Morrowmake/dsv41-flash-cmp170hx-recipe (1) · https://github.com/Morrowmake/qwen38-flash-next-cmp170hx-recipe (1) · https://github.com/mw00/project-maya#how-fast-is-it (1) · https://github.com/vektorprime/qwen38-flash-next-pp2 (1) · https://github.com/wgkbfcp/deepseek-v4.1-flash-1017-cmp170hx-recipe (1)