DeepSeek V4.1 Flash 1017 for the Lucebox engine (ROCMFP block formats)

Lucebox builds of DeepSeek's October 17 update of V4.1 Flash (deepseek-ai/DeepSeek-V4.1-Flash-1017, revision 1bae0e10) for the Lucebox DS4.1 engine (AMD Radeon AI PRO R9700 + Ryzen AI Max). Routed experts in the engine's ROCMFP block formats; attention, shared experts and head in Q8_0; both Engram tables embedded, read in place. Same recipes and the same quality harness as our September V4.1 Flash files, re-measured against the 1017 model.

For reference, the September files scored KL 0.266 (ROCMFP2S), 0.246 (MIX10) and 0.204 (fp2s with fp3 down everywhere) against the September model (teacher ppl 8.45) on the same tokens.

Which one: ROCMFP2S is what the 128 GB Lucebox installs (the hottest experts on the R9700, the rest on the Ryzen AI Max and the SSD). MIX10 is the better file where it fits resident (192 GB boards). ROCMFP23S is the best quality of the three, for two-box setups and boards with room for 179 GiB of experts. The -IM1017 files are the same recipes calibrated on the 1017 weights; prefer them (Lucebox Manage installs ROCMFP2S-IM1017).

Drafter

DeepSeek-V4.1-Flash-1017-DSpark-draft-MXFP4-Q8.gguf (7.4 GiB, sha256 in drafter/): the draft block DeepSeek ships inside the 1017 checkpoint (mtp.*, three stages, 128 routed experts top-3), routed experts repacked bit-exactly to MXFP4, dense projections Q8_0. Use it with these targets; a drafter from the September checkpoint proposes tokens the 1017 target accepts less often.

DeepSeek-V4.1-Flash-1017-DSpark-draft-Q2K-Q4K.gguf (4.6 GiB): the same drafter with Q2_K routed experts and Q4_K dense weights, for the 192 GB layout (--profile ds41-gorgon), where it leaves the R9700 room for larger prefill chunks.

Image input

DeepSeek-V4.1-Flash-1017-mmproj-BF16.gguf (0.90 GiB): the 1017 checkpoint's vision encoder (32-layer ViT and the aligner into the 5120-wide model), its 266 tensors copied byte for byte in BF16 by the engine's server/tools/export_ds4v_mmproj.py, with the checkpoint revision recorded. The 1017 update retrained the encoder with the language weights: pair it only with these 1017 targets (the engine refuses an encoder from another checkpoint). Add --mmproj DeepSeek-V4.1-Flash-1017-mmproj-BF16.gguf to the command below; images then go in as OpenAI image_url parts.

Run

luce_server DeepSeek-V4.1-Flash-1017-ROCMFP2S.gguf \
  --profile ds41-lucebox \
  --draft DeepSeek-V4.1-Flash-1017-DSpark-draft-MXFP4-Q8.gguf

With 192 GB of unified memory use --profile ds41-gorgon with the Q2K-Q4K drafter, where every expert stays resident. Lucebox Manage installs the files and picks the layout for the box. Engine and guide: https://github.com/Luce-Org/lucebox (server/docs/DS41.md).

Formats and measurement

Q2_0_ROCMFP2 (ggml type 107, 10 bytes per 32 weights, 2.5 bits, codebook {-1, 0, 1, 2} times a ue4m3 scale per 16 weights). fp2s files use type 107 with bit 7 of a half-block's scale byte meaning "mirror the codebook"; a reader that ignores that bit produces garbage, so use an engine build that honors it. Q3_0_ROCMFPX (type 104, 14 bytes per 32, 3.5 bits) on the down projections of the MIX10 file's first ten layers.

Every block's scale is chosen by an exhaustive search minimizing the activation-weighted squared error, weighted by an importance matrix (262,144 calibration tokens; the September V4.1 imatrix for the three plain files, an imatrix collected on the 1017 weights for the -IM1017 files: 196,608 tokens for ROCMFP2S-IM1017, the completed 262,144 tokens in imatrix/ for MIX10-IM1017 and ROCMFP23S-IM1017). KL is measured on 8,184 held-out tokens against the unquantized 1017 model through the same reference forward pass; eval/ holds the teacher log-probs and the tokens, results/ the raw scores and the file headers, imatrix/ the calibration statistics, so anyone can rescore a file. Toolkit: the Lucebox repo, lucebox_training/quantization.