Drilldown: Recipe

DeepSeek V4 Flash 0731 on 4x 8 GB (allover326 PP4 DSpark) (1) · DeepSeek V4.1 Flash on 4x 8 GB (GNmac) (1) · GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 layer split) (1) · GLM 5.3 Flash on 4x 8 GB (1) · GLM 5.3 on 8x 8 GB (vllm-170hx-glm5) (1) · Qwen3.8 27B on 1x 8 GB (W4A16 vLLM) (1) · Qwen3.8 Flash Next on 1x 8 GB (iIIusi0n SSD PLE) (1) · Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2) (1) · Qwen3.8 Flash Next on 3x 8 GB (clee PP3) (1)

DeepSeek-V4-Flash-0731 (1) · DeepSeek-V4.1-Flash-1017 (1) · GLM-5.3 (full 78-layer DSA MoE, not Flash) (1) · GLM-5.3-Flash (320B MoE, 18B active) (2) · Qwen3.8-27B (dense) (1) · Qwen3.8-Flash-Next (2) · Qwen3.8-Flash-Next (125B MoE per the recipe) (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 1.5.4 (built for SM80) with a small OpenAI-compatible server (1) · Morrowmake/vllm-cmp170hx (vLLM fork, container image) (1) · Morrowmake/vllm-cmp170hx branch ampere-glm53 plus fail-closed source patches (1) · Morrowmake/vllm-cmp170hx image (Qwen EXL3 release) with patch files mounted read-only (1) · vLLM (1) · vLLM a5a30471 plus a 13-file downstream patch (native Linux AIO PLE reader) (1) · vLLM nightly 7d0b4e57 plus the author's pure-Python fork clee/vllm-sm80 (branch q38fn/stack) (1) · vLLM with a pinned sm_80 patch stack (built locally from source) (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)

None (1) · 0.29.1rc1.dev402+ga5a30471f.ple1 (1) · base c3046d1 (recommended since 2026-08-13; first released on f8ea5bb); repo commit 3dd2d88 (1) · Flun/glm53-flash-cmp170hx-exl3 at 5933112 (1) · q38fn/stack on nightly-7d0b4e57 (1) · recipe 0.3.1; engine 18305948ac (1) · recipe 1.2.0 (1) · recipe 1.8.0, engine 819146023f (1) · vLLM 0.30.1rc1, fork commit 378c37b0 (1)

8 GB (9)

1 (2) · 2 (2) · 3 (1) · 4 (3) · 8 (1)

layer split (no tensor parallel), 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) · PP8 (also PP9 and PP10), TP1 (1) · single card (2) · TP2 with whole-expert expert parallel (1) · TP4 (default) or PP4 (1) · TP4 (default) or PP4 for x1/x4 links (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 (YaRN x4) (1) · 1,048,576 per request (1) · 262,144 (2) · 262,144 (up to 1,048,576) (1) · 262,144 advertised (tested through 32K) (1) · 262,144 native (YaRN to about 500k in the recipe's long variant) (1) · 392,960 (profile q8_384k) (1)

40 - 100 (4) · 100 - 500 (5)

None (2) · 1,800 - 2,600 (2) · 2,600 - 6,000 (5)

64 GB (1) · 64 GB, --gpu-memory-utilization 0.85 (1) · 64 GiB exposed, --gpu-memory-utilization 0.95 (1) · 64 GiB, --gpu-memory-utilization 0.96 (1) · 64 GiB, gpu-memory-utilization 0.92 (1) · 64 GiB, GPU_MEMORY_UTILIZATION 0.95 (1) · 64 GiB, GPU_UTIL 0.95 (fixed) (1) · 64 GiB, GPU_UTIL 0.96 at PP8 (1) · about 58 GiB of 64 GiB (116.6 GiB resident across both) (1)

None (3) · 170 W (author's 10-card run) (1) · 180 W (4) · 180 W (nvidia-smi -pl 180) (1)

None (5) · 580 or newer (measured on 610.57.04); open modules with HMM for the table-on-disk mode (1) · 610.43.03 (1) · CUDA 13 capable; measured on 610.43.02 (1) · NVIDIA 580 or newer; P2P needs the Morrowmake/cmpunlocker build (1)

None (4) · cmpunlocker (64 GiB) (1) · cmpunlocker (64 GiB, 74 SMs) with a P2P-capable build (1) · cmpunlocker (64 GiB, 74 SMs) with a P2P-capable driver for TP4 (1) · cmpunlocker (64 GiB, 74 SMs); Morrowmake/cmpunlocker for P2P (1) · cmpunlocker VRAM unlock (65,536 MiB per card) (1)

Reported · Proven · Other · None

68x0 (1) · c133 (1) · deejay9983 (1) · greencandle9753 (1) · iIIusi0n (1) · nathang6866 (1) · reala7096 (1) · voidgiraffe (1) · wakeupmbrc1 (1)

August 2026 (2) · September 2026 (2) · October 2026 (5)

https://github.com/allover326/deepseek-v4-cmp170hx (1) · https://github.com/arvkmr/vllm-170hx-glm5 (1) · https://github.com/clee/qwen3.8-flash-next-3x170hx (1) · https://github.com/ElectroHobo/qwen38-flash-next-w4a16-2x-cmp170hx-recipe (1) · https://github.com/Flun/glm53-flash-cmp170hx-exl3 (1) · https://github.com/GNmac-lab/deepseek-v41-flash-1017-cmp170hx (1) · https://github.com/iIIusi0n/qwen3.8-flash-next-cmp170hx (1) · https://github.com/Morrowmake/glm53-flash-cmp170hx-recipe (1) · https://www.reddit.com/r/LocalLLM/comments/1w2f03e/qwen3827b_w4a16_on_one_64_gb_cmp_170hx_16070/ (1)