Drilldown: Recipe

DeepSeek V4.1 Flash on 4x 8 GB (Morrowmake) (1) · DeepSeek V4.1 Flash on 6x 8 GB (Deepinfer-SM80 TP2 x PP3) (1) · GLM 5.3 Flash on 3x 8 GB (1) · GLM 5.3 Flash on 4x 8 GB (1)

DeepSeek-V4.1-Flash-1017 (2) · GLM-5.3-Flash (320B MoE, 18B active) (2)

Hybrid: routed experts EXL3 2.9 bpw (LibertAIDAI), everything else official FP8/BF16; KV BF16 (1) · Official checkpoint (FP8 dense, FP4 experts); software FP8 on sm_80 (1) · W4A16 (INT4 weights, group size 128, FP16 activations): canada-quant/GLM-5.3-Flash-W4A16-MTP (1) · W4A16, about 4.3 bpw, 163 GiB on disk (1)

Morrowmake/vllm-cmp170hx (via the glm53-flash-cmp170hx-recipe) (1) · Morrowmake/vllm-cmp170hx (vLLM fork, container image) (2) · vLLM PR #56201 head (79a7108d) plus the two Deepinfer-SM80 patches and 30 drop-in modules (1)

Deepinfer-SM80 2a828fc lineage; vLLM v0.28.1rc0 base (1) · not stated (1) · recipe 1.0.0, engine a0c57b5237 (1) · recipe 1.8.0, engine 819146023f (1)

8 GB (4)

3 (1) · 4 (2) · 6 (1)

PP3 (1) · TP2 x PP3, partition 14,14,12 (1) · TP4 (default) or PP4 (1) · TP4 with expert parallel (96 whole experts per card) (1)

1,048,576 (1) · 262,144 (2) · about 1,000,000 (8-bit KV cache) (1)

None (1) · 160.8 (1) · 287 (1) · 442.1 (1)

None (1) · 3,161 (1) · 3,818 (1) · 4,037 (1)

64 GB · 64 GiB · 64 GiB exposed, --gpu-memory-utilization 0.95 · 64 GiB, --gpu-memory-utilization 0.95 · 64 GB, --gpu-memory-utilization 0.94, TB_TORCH_MEM_CAP_GIB=61.5 · 64 GB, --gpu-memory-utilization 0.95 · 64 GB; weights 35.9 GB per rank (ranks 1-6), 37.6 GB rank 0, 46.3 GB rank 7 · 64 GiB (64,911 MiB usable), GPU_MEMORY_UTILIZATION 0.98 · 64 GiB, --gpu-memory-utilization 0.90 · 64 GiB, --gpu-memory-utilization 0.92 (miner co-resident) · 64 GiB, --gpu-memory-utilization 0.95 (fixed) · 64 GiB, --gpu-memory-utilization 0.96 · 64 GiB, --gpu-memory-utilization 0.97 · 64 GiB, gpu-memory-utilization 0.92 · 64 GiB, GPU_MEMORY_UTILIZATION 0.95 · 64 GiB, GPU_UTIL 0.95 (fixed) · 64 GiB, GPU_UTIL 0.96 at PP8 · 64 GiB; weights 45.5 GiB plus 10 GiB KV per card; GPU_UTIL 0.94 (fixed) · 64 GiB; weights PP0 39.48 GiB, PP1 45.53 GiB; --gpu-memory-utilization 0.9 · about 58 GiB of 64 GiB (116.6 GiB resident across both) · about 62.5 GB used of 64 GB, --gpu-memory-utilization 0.97 · about 62.9 / 62.3 GiB idle, peaks about 64.4 / 63.0 GiB with vision · gpu_split 58 GB per CMP, 70 GB on the PRO · Other · None

None (2) · 180 W (2)

None (2) · NVIDIA 580 or newer; P2P needs the Morrowmake/cmpunlocker build (2)

None (2) · cmpunlocker (64 GiB, 74 SMs); Morrowmake/cmpunlocker for P2P (2)

Proven (1) · Reported (3)

HalfVulpes (1) · shirogeek (1) · voidgiraffe (2)

October 16, 2026 (1) · October 19, 2026 (1) · October 24, 2026 (1) · October 28, 2026 (1)

https://discord.com/channels/1518634379604136178/1518695289278566450/1561743826945843304 (1) · https://github.com/HalfVulpes/Deepinfer-SM80 (1) · https://github.com/Morrowmake/dsv41-flash-cmp170hx-recipe (1) · https://github.com/Morrowmake/glm53-flash-cmp170hx-recipe (1)