View table: Recipe
More actions
Table structure:
- page - Page
- model - String
- quant - String
- engine - String
- engine_version - String
- cards - String (allowed values: 8 GB · 10 GB · both)
- card_count - Integer
- topology - String
- context - String
- decode_tps - Float
- prefill_tps - Float
- vram_per_card - String
- power_cap - String
- driver - String
- unlock_build - String
- flags - Text
- status - String (allowed values: Proven · Reported · Theory · Contested · Warning)
- tested_by - String
- tested_on - Date
- source - String
- files - Wikitext
This table has 30 rows altogether.
This table is defined by the template Recipe.
| _pageName | page | model | quant | engine | engine version | cards | card count | topology | context | decode tps | prefill tps | vram per card | power cap | driver | unlock build | flags | status | tested by | tested on | source | files |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| DeepSeek V4.1 Flash heretic on 4x 8 GB (edit) | DeepSeek V4.1 Flash heretic on 4x 8 GB | DeepSeek-V4.1-Flash-1017 heretic (pqhaz abliteration) | Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16 with the 38 heretic attn.wo_b tensors | Morrowmake/vllm-cmp170hx (vLLM fork, container image) | recipe 1.0.0 (commit d4d1f525), engine a0c57b52 | 8 GB | 4 | TP4 | 262,144 | 158.0 | 2,453.0 | 64 GiB, --gpu-memory-utilization 0.92 (miner co-resident) | EXTRA_ARGS="--gpu-memory-utilization 0.92", VLLM_ENGRAM_PLACEMENT=ssd, CUDA_DEVICE_ORDER=PCI_BUS_ID | Reported | darkmaniac7 | 2026-10-29 | https://github.com/Morrowmake/dsv41-flash-cmp170hx-recipe | ||||
| DeepSeek V4.1 Flash on 2x 8 GB (Project Maya) (edit) | DeepSeek V4.1 Flash on 2x 8 GB (Project Maya) | DeepSeek-V4.1-Flash-1017 (748B in the Maya README) | Maya GGUF (Maya-S 340 GB or Maya-XS 277 GB; Engram tables stay on SSD) | Project Maya (Strata and llama.cpp lineage) | 8 GB | 2 | engine-managed expert tiering across VRAM, host RAM and NVMe | 63.0 | 64 GB | Reported | pcsv. | 2026-10-25 | https://github.com/mw00/project-maya#how-fast-is-it | ||||||||
| DeepSeek V4.1 Flash on 4x 8 GB (alanoo fidelity build) (edit) | DeepSeek V4.1 Flash on 4x 8 GB (alanoo fidelity build) | DeepSeek-V4.1-Flash-1017 | 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 | GNmac-lab recipe image 1.2.0 (vLLM with sm_80 patches) plus three patches | GNmac-lab at 2ad5741ab1; image ds41-cmp170hx:1.2.0-fidelity | 8 GB | 4 | TP4, P2P on | 262,144 | 105.0 | 2,514.0 | 64 GiB (64,911 MiB usable), GPU_MEMORY_UTILIZATION 0.98 | 200 W | NVIDIA open 610.43.02 from a cmpunlocker fork with a P2P patch | cmpunlocker fork with P2P (without Morrowmake's TRAP31 patch) | LAYOUT=tp4, MAX_MODEL_LEN=262144, MAX_NUM_SEQS=8, GPU_MEMORY_UTILIZATION=0.98, P2P=1, SPEC_TOKENS=0 (DSpark off), ENGRAM_STORAGE=ssd from a tmpfs | Reported | alanoo | 2026-10-27 | https://github.com/alanoo81/deepseek-v41-flash-4x-cmp170hx | |
| DeepSeek V4.1 Flash on 4x 8 GB (CurtisAccelerate TC-MoE) (edit) | DeepSeek V4.1 Flash on 4x 8 GB (CurtisAccelerate TC-MoE) | DeepSeek-V4.1-Flash-1017 | 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 | Morrowmake/vllm-cmp170hx (bind-mounted sm80 package) plus the author's tensor-core MoE kernels | vllm-cmp170hx 18305948ac; TC-MoE fb9786ec | 8 GB | 4 | TP4, peer access off (NCCL_P2P_DISABLE=1), host all-reduce | 262,144 (weights address 1,048,576) | 144.2 | 1,523.0 | about 62.5 GB used of 64 GB, --gpu-memory-utilization 0.97 | 140 W | --quantization exl3 --tensor-parallel-size 4 --max-num-seqs 8 --max-num-batched-tokens 2048 --gpu-memory-utilization 0.97 --kv-cache-dtype fp8_ds_mla --engram-config cpu_offload --speculative-config dspark 5 | Reported | CurtisAccelerate | 2026-10-22 | https://github.com/CurtisAccelerate/DeepSeekv41-Flash-1017-EXL3-CMP170HX | |||
| DeepSeek V4.1 Flash on 4x 8 GB (GNmac) (edit) | DeepSeek V4.1 Flash on 4x 8 GB (GNmac) | DeepSeek-V4.1-Flash-1017 | 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 | vLLM with a pinned sm_80 patch stack (built locally from source) | recipe 1.2.0 | 8 GB | 4 | TP4 (default) or PP4 for x1/x4 links | 262,144 (up to 1,048,576) | 258.0 | 2,416.0 | 64 GiB, GPU_MEMORY_UTILIZATION 0.95 | 180 W (nvidia-smi -pl 180) | CUDA 13 capable; measured on 610.43.02 | cmpunlocker (64 GiB, 74 SMs) with a P2P-capable driver for TP4 | LAYOUT=tp4, MAX_MODEL_LEN 262144, MAX_NUM_SEQS 8, GPU_MEMORY_UTILIZATION 0.95, P2P 1, DSV41_GATE_TAU 0.10, DSV41_DYNK accept:3,5 | Proven | greencandle9753 | 2026-10-26 | https://github.com/GNmac-lab/deepseek-v41-flash-1017-cmp170hx | |
| DeepSeek V4.1 Flash on 4x 8 GB (Morrowmake) (edit) | DeepSeek V4.1 Flash on 4x 8 GB (Morrowmake) | DeepSeek-V4.1-Flash-1017 | Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16; KV BF16 | Morrowmake/vllm-cmp170hx (vLLM fork, container image) | recipe 1.0.0, engine a0c57b5237 | 8 GB | 4 | TP4 with expert parallel (96 whole experts per card) | 262,144 | 287.0 | 3,818.0 | 64 GiB exposed, --gpu-memory-utilization 0.95 | 180 W | NVIDIA 580 or newer; P2P needs the Morrowmake/cmpunlocker build | cmpunlocker (64 GiB, 74 SMs); Morrowmake/cmpunlocker for P2P | P2P=auto, VLLM_ENGRAM_PLACEMENT=auto (ram or ssd), MAX_SEQS 8, MAX_BATCHED 2048, DSpark drafter up to 7 tokens | Reported | voidgiraffe | 2026-10-24 | https://github.com/Morrowmake/dsv41-flash-cmp170hx-recipe | |
| DeepSeek V4.1 Flash on 4x 8 GB plus RTX PRO 6000 (ExLlamaV3 TP5) (edit) | DeepSeek V4.1 Flash on 4x 8 GB plus RTX PRO 6000 (ExLlamaV3 TP5) | DeepSeek-V4.1-Flash-1017 | 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 | ExLlamaV3 fork (anoane/exllamav3-anemone) + TabbyAPI fork (anoane/tabbyAPI) | tag recipe-09 (309ef6d1); TabbyAPI deepseek-v41 (ce250ff) | 8 GB | 4 | TP5 across four CMP 170HX and one RTX PRO 6000 Blackwell (native backend, reduces through host memory); MTP draft on the PRO | 1,048,576 FP16 | 99.9 | 2,116.0 | gpu_split 58 GB per CMP, 70 GB on the PRO | 150 W per CMP (PRO 400 W) | NVIDIA open 610.57.04 patched with Morrowmake/cmpunlocker 1f2c498 (profile 8gb) | Morrowmake/cmpunlocker 1f2c498; modprobe RmForceEnableGen2=1, RMPcieLinkSpeed=0x1, ForceP2P=0x11 | '"`UNIQ--nowiki-0000004E-QINU`"' | Reported | anoane | 2026-10-20 | https://github.com/anoane/09-deepseek-v41-flash-1017-exl3-tp5-1m | |
| DeepSeek V4.1 Flash on 6x 8 GB (Deepinfer-SM80 TP2 x PP3) (edit) | DeepSeek V4.1 Flash on 6x 8 GB (Deepinfer-SM80 TP2 x PP3) | DeepSeek-V4.1-Flash-1017 | Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 | vLLM PR #56201 head (79a7108d) plus the two Deepinfer-SM80 patches and 30 drop-in modules | Deepinfer-SM80 2a828fc lineage; vLLM v0.28.1rc0 base | 8 GB | 6 | TP2 x PP3, partition 14,14,12 | 1,048,576 | 160.8 | 4,037.0 | 64 GiB | VLLM_PP_LAYER_PARTITION=14,14,12, --speculative-config dspark num_speculative_tokens 5, --max-model-len 1048576 | Reported | HalfVulpes | 2026-10-28 | https://github.com/HalfVulpes/Deepinfer-SM80 | ||||
| DeepSeek V4.1 Flash on 6x 8 GB (TP6) (edit) | DeepSeek V4.1 Flash on 6x 8 GB (TP6) | DeepSeek-V4.1-Flash-1017 | Official checkpoint: dense MXFP8 through an SM80 Marlin W8A16 path, routed experts MXFP4 through W4A16; KV FP8 | Custom SM80 vLLM (patched from public commit 79a7108d9) | reports 0.0.0; patches/public-base-to-runtime.patch (109 files) | 8 GB | 6 | TP6 (virtual-head padding), PP1 | 1,048,576 | 78.55 | 64 GB, --gpu-memory-utilization 0.95 | --tensor-parallel-size 6 --kv-cache-dtype fp8 --block-size 128 --max-num-seqs 16 --max-num-batched-tokens 2048 --engram-config storage disk --speculative-config dspark 5; --hf-overrides virtual heads 96 / o_groups 12 | Reported | wgkbfcp | 2026-10-19 | https://github.com/wgkbfcp/deepseek-v4.1-flash-1017-cmp170hx-recipe | |||||
| DeepSeek V4.1 Flash on 8x 8 GB (Deepinfer-SM80 TP4 x PP2) (edit) | DeepSeek V4.1 Flash on 8x 8 GB (Deepinfer-SM80 TP4 x PP2) | DeepSeek-V4.1-Flash-1017 | Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 | vLLM PR #56201 head (79a7108d) plus the two Deepinfer-SM80 patches and 30 drop-in modules | Deepinfer-SM80, vLLM v0.28.1rc0 base | 8 GB | 8 | TP4 x PP2, partition 20,20 | 1,048,576 | 153.35 | 3,463.0 | 64 GiB, --gpu-memory-utilization 0.95 | VLLM_PP_LAYER_PARTITION=20,20, --gpu-memory-utilization 0.95, max_num_seqs 12, DSpark K=5 | Reported | HalfVulpes | 2026-10-28 | https://github.com/HalfVulpes/Deepinfer-SM80 | ||||
| DeepSeek V4.1 Flash on 8x 8 GB (Zanooda PP8) (edit) | DeepSeek V4.1 Flash on 8x 8 GB (Zanooda PP8) | DeepSeek-V4.1-Flash-1017 | Official checkpoint as downloaded: MXFP4 routed experts as Marlin W4A16, MXFP8 dense as Marlin W8A16; KV fp8 | vLLM PR #56201 head (79a7108d) plus a 20-patch sm_80 series; image zanooda/vllm-sm80-ds41f:v41-sm80 | recipe 31318fd; patches 0001-0020 | 8 GB | 8 | PP8, partition 5,5,5,5,5,5,5,5 | 1,048,576 | 117.0 | 6,066.0 | 64 GB; weights 35.9 GB per rank (ranks 1-6), 37.6 GB rank 0, 46.3 GB rank 7 | 180 W (one card unstable at 250 W) | cmpunlocker with the BAR1 P2P patches (64 GiB BAR1 per card) | --pipeline-parallel-size 8 --max-model-len 1048576 --gpu-memory-utilization 0.95 --block-size 128 --kv-cache-dtype fp8 --max-num-batched-tokens 4096 --max-num-seqs 32; DSpark 5; VLLM_DSV41_SPREAD_DECODE=2; --engram-config storage cpu | Reported | Zanooda | 2026-10-19 | https://github.com/Zanooda/deepseek-v4.1-flash-cmp170hx | ||
| GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 layer split) (edit) | GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 layer split) | GLM-5.3-Flash (320B MoE, 18B active) | EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3 | ExLlamaV3 1.5.4 (built for SM80) with a small OpenAI-compatible server | Flun/glm53-flash-cmp170hx-exl3 at 5933112 | 8 GB | 2 | layer split (no tensor parallel), no P2P | 392,960 (profile q8_384k) | 60.0 | about 58 GiB of 64 GiB (116.6 GiB resident across both) | scripts/serve.sh --profile q8_384k (-cs 393216 -cq 8 -gs 62,63 --q8-staging --q8-head-gpu0 -chunk_size 2048) | Proven | 68x0 | 2026-10-20 | https://github.com/Flun/glm53-flash-cmp170hx-exl3 | |||||
| GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) (edit) | GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) | GLM-5.3-Flash (320B MoE, 18B active) | EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3; DFlash2 drafter converted to EXL3 6 bpw | ExLlamaV3 (patched, pinned) | upstream dev 0f685aa9fd plus 9 patches; repo db01c256 | 8 GB | 2 | TP2 over NCCL (GPU P2P), DFlash2 K7 plus hybrid drafting | 392,960 per request; 655,360-token Q8 KV pool shared by two requests | 107.67 | 1,533.0 | about 62.9 / 62.3 GiB idle, peaks about 64.4 / 63.0 GiB with vision | 180 W (author's cards); about 170 W with a modest memory OC in his Discord numbers | 610.57.04 | cmpunlocker (64 GiB); P2P-capable driver for the NCCL backend | scripts/run_tp.sh: -cq 8 -cs 655360 --context-cap 392960 -ndt 7 -chunk_size 2048 -ambs 2 -gs 62,63 -tp --experimental-dflash2-tp -tpb nccl --q8-staging --q8-head-gpu0 --concurrent-max-sequences 2; EXL3_MLA_KIDX_RING=32 | Reported | nestaa51 | 2026-10-13 | https://github.com/alexvnesta/glm53-flash-cmp170hx-tp | |
| GLM 5.3 Flash on 2x 8 GB (Project Maya) (edit) | GLM 5.3 Flash on 2x 8 GB (Project Maya) | GLM-5.3-Flash (321B in the Maya README) | Maya-M GGUF (116 GB), every expert in VRAM; Maya-S (96.5 GB) and Maya-S24 (94.7 GB) also reported | Project Maya (Strata and llama.cpp lineage, EXL3 loading, DFlash2 drafter) | v1.0.6 (this row) | 8 GB | 2 | engine-managed split with expert tiering across VRAM, RAM and NVMe | 61.3 | 514.0 | 64 GB | ./maya.sh --bench (three greedy 256-token answers for decode, an 8K-token prompt for prefill) | Reported | ZackO2o | https://github.com/mw00/project-maya#how-fast-is-it | ||||||
| GLM 5.3 Flash on 3x 8 GB (edit) | GLM 5.3 Flash on 3x 8 GB | GLM-5.3-Flash (320B MoE, 18B active) | W4A16, about 4.3 bpw, 163 GiB on disk | Morrowmake/vllm-cmp170hx (via the glm53-flash-cmp170hx-recipe) | not stated | 8 GB | 3 | PP3 | about 1,000,000 (8-bit KV cache) | 64 GiB | Morrowmake's recipe with pipeline parallel 3; KV cache quantized to 8 bits | Reported | shirogeek | 2026-10-19 | https://discord.com/channels/1518634379604136178/1518695289278566450/1561743826945843304 | ||||||
| GLM 5.3 Flash on 4x 8 GB (edit) | GLM 5.3 Flash on 4x 8 GB | GLM-5.3-Flash (320B MoE, 18B active) | W4A16 (INT4 weights, group size 128, FP16 activations): canada-quant/GLM-5.3-Flash-W4A16-MTP | Morrowmake/vllm-cmp170hx (vLLM fork, container image) | recipe 1.8.0, engine 819146023f | 8 GB | 4 | TP4 (default) or PP4 | 262,144 | 442.1 | 3,161.0 | 64 GiB exposed, --gpu-memory-utilization 0.95 | 180 W | NVIDIA 580 or newer; P2P needs the Morrowmake/cmpunlocker build | cmpunlocker (64 GiB, 74 SMs); Morrowmake/cmpunlocker for P2P | LAYOUT=tp4 or LAYOUT=pp4, P2P=auto, --max-num-seqs 8, --gpu-memory-utilization 0.95, --max-num-batched-tokens 3463 (TP4) or 2312 (PP4) | Proven | voidgiraffe | 2026-10-16 | https://github.com/Morrowmake/glm53-flash-cmp170hx-recipe | |
| GLM 5.3 on 8x 8 GB (Int4-Int8Mix PP8) (edit) | GLM 5.3 on 8x 8 GB (Int4-Int8Mix PP8) | GLM-5.3 (full 78-layer, not Flash) | GLM-5.3-Int4-Int8Mix (INT4 / INT8 mix); KV fp8_ds_mla | vLLM (build not stated) | 8 GB | 8 | PP8, TP1 | 1,048,576 | 25.0 | 64 GiB, --gpu-memory-utilization 0.97 | 200 W | --pipeline-parallel-size 8 --tensor-parallel-size 1 --speculative-config mtp 5 --attention-backend TRITON_MLA_SPARSE --kv-cache-dtype fp8_ds_mla --block-size 64 --no-async-scheduling | Reported | 1a2b3c0967 | 2026-10-07 | https://discord.com/channels/1518634379604136178/1528826330454429706/1557523316120748073 | |||||
| GLM 5.3 on 8x 8 GB (vllm-170hx-glm5) (edit) | GLM 5.3 on 8x 8 GB (vllm-170hx-glm5) | GLM-5.3 (full 78-layer DSA MoE, not Flash) | GLM-5.3-Int4-Int8Mix-AWQ-g64; drafter incoai/GLM-5.3-DFlash2 | Morrowmake/vllm-cmp170hx branch ampere-glm53 plus fail-closed source patches | vLLM 0.30.1rc1, fork commit 378c37b0 | 8 GB | 8 | PP8 (also PP9 and PP10), TP1 | 1,048,576 per request | 48.7 | 1,833.0 | 64 GiB, GPU_UTIL 0.96 at PP8 | 170 W (author's 10-card run) | --attention-backend TRITON_MLA_SPARSE, packed fp8_ds_mla KV, --block-size 128, async scheduling required; PP8 partition 11,10,10,10,10,10,9,8 | Proven | reala7096 | 2026-09-29 | https://github.com/arvkmr/vllm-170hx-glm5 | |||
| MiMo V2.6 Flash on 4x 8 GB (edit) | MiMo V2.6 Flash on 4x 8 GB | MiMo-V2.6-Flash-RL (309B MoE, 15B active; text, image, video and audio input) | NVFP4 (ModelOpt mixed precision): ProCreations/MiMo-V2.6-Flash-RL-NVFP4, run as 4-bit weights with BF16 activations through Marlin | Morrowmake/vllm-cmp170hx image (tag qwen38-recipe-1.0.0) plus one mounted file (vLLM #58235 vision-sink fix) | recipe 0.2.1; engine 18305948ac | 8 GB | 4 | TP4 (P2P on); expert parallel off | 1,048,576 | 274.9 | 2,722.0 | 64 GiB; weights 45.5 GiB plus 10 GiB KV per card; GPU_UTIL 0.94 (fixed) | 180 W | 580 or newer (measured on 610.57.04) | cmpunlocker (64 GiB, 74 SMs) with a P2P-capable build | --tensor-parallel-size 4 --kv-cache-dtype bfloat16 --block-size 128 --max-model-len 1048576 --gpu-memory-utilization 0.94 --kv-cache-memory 10737418240 --max-num-seqs 32 --max-num-batched-tokens 2048; DFlash 7/5/3 by load | Reported | nathang6866 | 2026-10-18 | https://github.com/ElectroHobo/mimo-v26-flash-cmp170hx-recipe | |
| Qwen3.8 27B on 1x 8 GB (W4A16 vLLM) (edit) | Qwen3.8 27B on 1x 8 GB (W4A16 vLLM) | Qwen3.8-27B (dense) | W4A16 (INT4 weights, Marlin kernels) | vLLM | 8 GB | 1 | single card | 262,144 native (YaRN to about 500k in the recipe's long variant) | 170.0 | 64 GB | Reported | kellertreppe123 | 2026-10-05 | https://www.reddit.com/r/LocalLLM/comments/1w2f03e/qwen3827b_w4a16_on_one_64_gb_cmp_170hx_16070/ | |||||||
| Qwen3.8 Flash Next on 1x 8 GB (ElectroHobo 3 bpw) (edit) | Qwen3.8 Flash Next on 1x 8 GB (ElectroHobo 3 bpw) | Qwen3.8-Flash-Next (125B MoE per the recipe) | AutoRound mixed 2/3-bit (klee100, 2.998 effective bpw, 47 GiB on card) plus halt95's FP8 n-gram table in host RAM | Morrowmake/vllm-cmp170hx image with patch files mounted read-only | recipe 0.2.1; engine 18305948ac | 8 GB | 1 | single card | 262,144 | 296.4 | 4,243.0 | 64 GiB, --gpu-memory-utilization 0.95 (fixed) | 180 W | 580 or newer (measured on 610.57.04) | cmpunlocker (64 GiB, 74 SMs) | '"`UNIQ--nowiki-0000002D-QINU`"' | Reported | nathang6866 | 2026-10-23 | https://github.com/ElectroHobo/qwen38-flash-next-3bpw-1x-cmp170hx-recipe | |
| Qwen3.8 Flash Next on 1x 8 GB (iIIusi0n SSD PLE) (edit) | Qwen3.8 Flash Next on 1x 8 GB (iIIusi0n SSD PLE) | Qwen3.8-Flash-Next | AutoRound 3 bpw + MTP: klee100/Qwen3.8-Flash-Next-AutoRound-3bpw-MTP; 95.37 GiB BF16 PLE table on SSD | vLLM a5a30471 plus a 13-file downstream patch (native Linux AIO PLE reader) | 0.29.1rc1.dev402+ga5a30471f.ple1 | 8 GB | 1 | single card | 262,144 advertised (tested through 32K) | 110.99 | 2,636.0 | 64 GiB, --gpu-memory-utilization 0.96 | 180 W | 610.43.03 | cmpunlocker (64 GiB) | --max-num-seqs 16 --max-num-batched-tokens 2048 --gpu-memory-utilization 0.96 --mamba-cache-mode align --enable-prefix-caching cudagraph FULL_AND_PIECEWISE; MTP 1; ple_ssd_offload true, ple_ssd_io_depth 256, ple_ssd_cache_mb 512 | Proven | iIIusi0n | 2026-09-19 | https://github.com/iIIusi0n/qwen3.8-flash-next-cmp170hx | |
| Qwen3.8 Flash Next on 2x 8 GB (DUFRelic PP2) (edit) | Qwen3.8 Flash Next on 2x 8 GB (DUFRelic PP2) | Qwen3.8-Flash-Next | W4A16 INT4 g128 with the PLE table repacked INT4 g32: zankich/Qwen3.8-Flash-Next-W4A16-Merlin-INT4PLE-FP8KV; KV BF16 | vLLM v0.30.0 + zankich INT4-PLE plugin + six hash-guarded patch layers | vLLM v0.30.0 (ced6857a); plugin zankich/vllm v0.29.0z (6609bfef) | 8 GB | 2 | PP2, MTP depth 3 | 262,144 | 133.39 | 6,213.66 | 64 GiB, --gpu-memory-utilization 0.90 | 140 W (200 W observed-draw abort guard) | NVIDIA 610.43.02 | amoghmunikote/cmpunlocker 6c442eeb (patch 0016 SM reconfig, 74 SMs) plus Bayley P2P patches 0011/0013/0015 | '"`UNIQ--nowiki-0000003D-QINU`"' | Reported | dufrelic | 2026-10-03 | https://github.com/DUFRelic/qwen38-flash-next-recipe | |
| Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2) (edit) | Qwen3.8 Flash Next on 2x 8 GB (ElectroHobo W4A16 TP2) | Qwen3.8-Flash-Next (125B MoE per the recipe) | W4A16: halt95/Qwen3.8-Flash-Next-W4A16-Merlin (AutoRound INT4 g128 experts, INT8 GDN projections, FP8 n-gram table) | Morrowmake/vllm-cmp170hx image (Qwen EXL3 release) with patch files mounted read-only | recipe 0.3.1; engine 18305948ac | 8 GB | 2 | TP2 with whole-expert expert parallel | 262,144 | 349.7 | 4,696.0 | 64 GiB, GPU_UTIL 0.95 (fixed) | 180 W | 580 or newer (measured on 610.57.04); open modules with HMM for the table-on-disk mode | cmpunlocker (64 GiB, 74 SMs) with a P2P-capable build | --tensor-parallel-size 2 --enable-expert-parallel --kv-cache-dtype bfloat16 --engram-config cpu_offload --max-num-seqs 128 --block-size 912 cudagraph FULL_AND_PIECEWISE; DFlash 2 drafter 7/5/3/2 by load | Proven | nathang6866 | 2026-10-23 | https://github.com/ElectroHobo/qwen38-flash-next-w4a16-2x-cmp170hx-recipe | |
| Qwen3.8 Flash Next on 2x 8 GB (mannydelrio PP2) (edit) | Qwen3.8 Flash Next on 2x 8 GB (mannydelrio PP2) | Qwen3.8-Flash-Next-Uncensored (refusal-removed, from orcarouter) | AWQ W4A16 g32 (leoncca/Qwen3.8-Flash-Next-Uncensored-AWQ-g32); PLE table converted to INT4 and served from SSD | wtdcode/vllm-backport 3142ded1 plus a two-file patch and two tinybeast PLE plugins (venv, no container) | vllm-backport 3142ded1 | 8 GB | 2 | PP2 (layer partition 25,23), no P2P | 262,144 | 177.8 | 7,006.0 | 64 GB, --gpu-memory-utilization 0.94, TB_TORCH_MEM_CAP_GIB=61.5 | 180 W | nvidia-open 610.43.02 + cmpunlocker v0.5 or newer | cmpunlocker v0.5+ (sudo ./install.sh --profile=8gb --no-iommu) | '"`UNIQ--nowiki-0000004A-QINU`"' | Reported | mrmannyatx | 2026-10-15 | https://github.com/mannydelrio/qwen38-flash-next-2x-cmp170hx | |
| Qwen3.8 Flash Next on 2x 8 GB (Morrowmake EXL3) (edit) | Qwen3.8 Flash Next on 2x 8 GB (Morrowmake EXL3) | Qwen3.8-Flash-Next | EXL3 5.05 bpw; PLE (n-gram) table read from SSD | Morrowmake/vllm-cmp170hx (EXL3 path) | tag qwen38-recipe-1.0.0, engine 18305948ac | 8 GB | 2 | TP2 with expert parallel | 262,144 | 152.6 | 2,651.0 | 64 GiB, --gpu-memory-utilization 0.95 | --quantization exl3 --load-format exl3 --tensor-parallel-size 2 --enable-expert-parallel --engram-config cpu_offload false, VLLM_EXL3_PLE_PLACEMENT=ssd, cudagraph PIECEWISE, MTP 3 | Reported | voidgiraffe | 2026-10-16 | https://github.com/Morrowmake/qwen38-flash-next-cmp170hx-recipe | ||||
| Qwen3.8 Flash Next on 2x 8 GB (Plean PP2) (edit) | Qwen3.8 Flash Next on 2x 8 GB (Plean PP2) | Qwen3.8-Flash-Next | Mixed INT4 AutoRound (Minachist/Qwen3.8-Flash-Next-INT4-Mixed-AutoRound, about 65 GB) with the full BF16 PLE table on NVMe | vLLM 0.29.1rc1 nightly plus the checkpoint author's patches | 8 GB | 2 | PP2, MTP 3, no P2P | 262,144 (256K in the post) | 197.0 | 5,800.0 | 64 GB | --pipeline-parallel-size 2, MTP 3 draft tokens, max concurrency 6, BF16 PLE table offloaded to NVMe | Reported | plean. | 2026-10-07 | https://www.bozocentral.com/2x-cmp170hx-qwen3-flash-next | |||||
| Qwen3.8 Flash Next on 2x 8 GB (vektorprime PP2) (edit) | Qwen3.8 Flash Next on 2x 8 GB (vektorprime PP2) | Qwen3.8-Flash-Next | AWQ INT4: cyankiwi/Qwen3.8-Flash-Next-AWQ-INT4; 95 GiB BF16 PLE table in host RAM (FP8 table optional) | vLLM nightly + PR #53899 (a5530b9) + the repo's patch scripts | 0.1.dev1+ga5530b90c.d20260829 | 8 GB | 2 | PP2 with PLE CPU offload | 262,144 | 75.4 | 64 GiB; weights PP0 39.48 GiB, PP1 45.53 GiB; --gpu-memory-utilization 0.9 | --pipeline-parallel-size 2 --max-num-seqs 8 --mamba-cache-mode align --mamba-ssm-cache-dtype float32 --speculative-config mtp --enable-prefix-caching --async-scheduling; VLLM_PLE_CPU_OFFLOAD=1, VLLM_GDN_DECODE_KERNEL=triton, PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True | Reported | vektor_prime | 2026-08-29 | https://github.com/vektorprime/qwen38-flash-next-pp2 | |||||
| Qwen3.8 Flash Next on 3x 8 GB (clee PP3) (edit) | Qwen3.8 Flash Next on 3x 8 GB (clee PP3) | Qwen3.8-Flash-Next | 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) | vLLM nightly 7d0b4e57 plus the author's pure-Python fork clee/vllm-sm80 (branch q38fn/stack) | q38fn/stack on nightly-7d0b4e57 | 8 GB | 3 | PP3, no P2P (works at PP2 on two cards) | 1,048,576 (YaRN x4) | 74.4 | 4,831.0 | 64 GiB, gpu-memory-utilization 0.92 | pipeline-parallel-size 3, block-size 1728, kv-cache-dtype fp8_e4m3, max-model-len 1048576, max-num-seqs 8, max-num-batched-tokens 8192, FULL_AND_PIECEWISE; NCCL_P2P_DISABLE=1, VLLM_PP_LAYER_PARTITION=16,16,16, VLLM_PLE_STAGED=1 | Proven | c133 | 2026-10-10 | https://github.com/clee/qwen3.8-flash-next-3x170hx | ||||
| Qwen3.8 Flash Next on 4x 8 GB (edit) | Qwen3.8 Flash Next on 4x 8 GB | Qwen3.8-Flash-Next | AWQ W4A16: wtdcode/Qwen3.8-Flash-Next-AWQ-W4A16 | wtdcode/vllm-backport (image lazymio/vllm-backport:latest-sm80) | latest-sm80 tag as of 2026-10-16 | 8 GB | 4 | TP4 with expert parallel, MTP 3 | 1,000,000 (YaRN x4 from 262,144) | 116.49 | 2,101.4 | 64 GB | --tensor-parallel-size 4 --enable-expert-parallel --speculative-config mtp 3 --max-model-len 1000000 --max-num-seqs 8 --max-num-batched-tokens 2048 cudagraph FULL_AND_PIECEWISE --prefix-match-unit 16 --mamba-cache-mode align; VLLM_PLE_CPU_OFFLOAD=0 | Reported | matstatman | 2026-10-16 | https://discord.com/channels/1518634379604136178/1528826330454429706/1560602001908170753 |