DeepSeek-V4-Flash-0731 for Strix Halo

82/92 quality in one 98.29 GB GGUF file.

Built for 128 GB AMD Strix Halo systems. It loads unsplit on the Radeon 8060S iGPU and needs no sidecar files.

Quality

The full score matches the published reference. This file averages 2.766 bits per model weight, about 4% less than the 2.88-bit reference.

None of the 92 test questions were used while preparing this file. The published reference used 75 of them during its own preparation. Both results used the same grader and reasoning allowance.

Quality or speed

Quality mode is the recommended setting. Faster mode is 23% quicker but misses one additional COMPSEC question. The full 82/92 evaluation was run only in quality mode.

The current DSpark helper model makes this file slower overall, so it is not recommended yet.

Download

DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf

• Size: 98,294,917,184 bytes
• One file, no sidecars
• SHA-256: 7c0789d190fdd2acad93255825822ca276f29d13f9410f2ac65f5f7a542b0a38

Run

dflash_server DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf \
  --target-device hip:0 \
  --max-ctx 8192

Use the normal automatic memory settings on Strix Halo. If the machine also has a discrete GPU, expose only the iGPU with HIP_VISIBLE_DEVICES.

Until support reaches the main dflash release, use the feat/qtype106-down-surface branch of GeometricAGI/lucebox-hub.

Artifact and evaluation by Geometric-AI. Mirrored byte-for-byte by Lucebox.

Vision: DeepSeek-V4-Flash-Vision-Exp

The same Strix Halo format for deepseek-ai/DeepSeek-V4-Flash-Vision-Exp, which reads images as well as text. It is a separate model, not a patch on the file above: every decoder weight differs from 0731, and it carries the image router biases.

Built with the recipe of the 0731 file above: routed gate and up experts in fp2, down experts in fp2 on the same 15 layers and fp3 elsewhere, dense projections in ROCmFP4, the token embedding in Q6_K, codebooks inside the file. Calibrated with the published llama.cpp importance matrix for this checkpoint (812 chunks of 512 tokens), applied expert by expert.

Quality

Measured on one Strix Halo, all six routed experts.

The reference is ggml-org's MXFP4 conversion, which holds the model's native FP4 experts exactly. Image answers on the sanity set and on requests with one to four images were all correct. The 92-question text evaluation of the 0731 file has not been run on this one yet.

Run

Image input needs the projector and a build that includes Luce-Org/lucebox#722 until it reaches the main release.

LUCE_DS4_SPEC=1 \
LUCE_DS4_DRAFT=/path/to/DeepSeek-V4-Flash-0731-DSpark-draft-Q4RMFP4-denseF16.gguf \
LUCE_DS4_SPARSE_DECODE_FLASH=1 \
luce_server DeepSeek-V4-Flash-Vision-Exp-ROCMFPX-MIX-STRIX.gguf \
  --mmproj DeepSeek-V4-Flash-Vision-Exp-mmproj-BF16.gguf \
  --target-device hip:0 --max-ctx 131072 --chunk 8192 \
  --cache-type-k q4_0 --cache-type-v q4_0 \
  --ds4-fused-decode --ds4-fused-verify-f16-kv \
  --ds4-expert-top-k 6 --ds4-prefill sparse

With the DSpark drafter from Lucebox/DeepSeek-V4-Flash-0731-DSpark-GGUF, text decodes at 25 to 37 tok/s on 256-token answers (30 on average). Image requests decode without the drafter, at about 22 tok/s. Send images as base64 JPEG or PNG in OpenAI chat-completion image_url parts, up to four per request.

Vision artifact and evaluation by Lucebox.