Drilldown: Recipe

DeepSeek V4.1 Flash on 4x 8 GB (CurtisAccelerate TC-MoE) (1) · GLM 5.3 Flash on 2x 8 GB (ExLlamaV3 TP2) (1)

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

EXL3 3.05 bpw: turboderp/GLM-5.3-Flash-exl3; DFlash2 drafter converted to EXL3 6 bpw (1) · 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 (1)

ExLlamaV3 (patched, pinned) (1) · Morrowmake/vllm-cmp170hx (bind-mounted sm80 package) plus the author's tensor-core MoE kernels (1)

upstream dev 0f685aa9fd plus 9 patches; repo db01c256 (1) · vllm-cmp170hx 18305948ac; TC-MoE fb9786ec (1)

8 GB · Other · None

2 (1) · 4 (1)

TP2 over NCCL (GPU P2P), DFlash2 K7 plus hybrid drafting (1) · TP4, peer access off (NCCL_P2P_DISABLE=1), host all-reduce (1)

262,144 (weights address 1,048,576) (1) · 392,960 per request; 655,360-token Q8 KV pool shared by two requests (1)

107.67 (1) · 144.2 (1)

1,523 (1) · 1,533 (1)

about 62.5 GB used of 64 GB, --gpu-memory-utilization 0.97 (1) · about 62.9 / 62.3 GiB idle, peaks about 64.4 / 63.0 GiB with vision (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 (1) · 610.57.04 (1)

None (1) · cmpunlocker (64 GiB); P2P-capable driver for the NCCL backend (1)

Reported (2)

CurtisAccelerate (1) · nestaa51 (1)

October 13, 2026 (1) · October 22, 2026 (1)

https://github.com/alexvnesta/glm53-flash-cmp170hx-tp (1) · https://github.com/CurtisAccelerate/DeepSeekv41-Flash-1017-EXL3-CMP170HX (1)