Drilldown: Recipe

DeepSeek V4.1 Flash on 4x 8 GB (CurtisAccelerate TC-MoE) (1)

DeepSeek-V4.1-Flash-1017 (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)

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

vllm-cmp170hx 18305948ac; TC-MoE fb9786ec (1)

8 GB · Other · None

4 (1)

TP4, peer access off (NCCL_P2P_DISABLE=1), host all-reduce (1)

262,144 · 1,048,576 · None · 1,000,000 (YaRN x4 from 262,144) · 1,048,576 (YaRN x4) · 1,048,576 FP16 · 1,048,576 per request · 262,144 (256K in the post) · 262,144 (up to 1,048,576) · 262,144 (weights address 1,048,576) · 262,144 advertised (tested through 32K) · 262,144 native (YaRN to about 500k in the recipe's long variant) · 392,960 (profile q8_384k) · 392,960 per request; 655,360-token Q8 KV pool shared by two requests · about 1,000,000 (8-bit KV cache) · Other

144.2 (1)

1,523 (1)

about 62.5 GB used of 64 GB, --gpu-memory-utilization 0.97 (1)

140 W (1)

None (1)

None (1)

Reported (1)

CurtisAccelerate (1)

October 22, 2026 (1)

https://github.com/CurtisAccelerate/DeepSeekv41-Flash-1017-EXL3-CMP170HX (1)