Lucebox × Geometric: DeepSeek V4 Flash 0731 reaches 32.7 tok/s on AMD Strix Halo | lucebox

August 2026<br>By Lucebox, with Geometric<br>Lucebox × Geometric: DeepSeek V4 Flash 0731 reaches 32.7 tok/s on AMD Strix Halo<br>Lucebox Engine now serves DeepSeek V4 Flash 0731 from one 98.29 GB GGUF on a 128 GB AMD Strix Halo system. The artifact scores 82/92 on ds4-eval-92 and reaches 32.7 tok/s with DSpark. Across our fixed HumanEval, GSM8K, and MATH evaluation, it averages 27.9 tok/s.<br>We added support for DeepSeek-V4-Flash-0731 to Lucebox Engine together with Geometric. The collaboration covers the adaptive ROCmFPX formats, HIP and CUDA kernels, embedded codebook loading, runtime validation, and the monolithic device load needed by this artifact. Geometric PR #28 records the adaptive-quantization contribution; Lucebox PR #593 carries that work into Lucebox Engine.<br>Geometric builds tools that automatically discover and verify CUDA and Triton kernels, then ships the resulting speedups as GitHub pull requests. Their work is aimed at teams deploying machine learning across GPU fleets.<br>The supported target is available from Lucebox on Hugging Face. It is a single 98.29 GB GGUF for a 128 GB AMD Ryzen AI MAX+ 395 / Radeon 8060S configuration. Per-expert quantization data, codebooks, and runtime metadata are embedded in the same file.<br>Quality benchmark<br>The release validation, documented on the Lucebox model card, ran the complete 92-item ds4-eval suite with a 15,488-token thinking budget, matching the published reference. The supported artifact scored 82/92 while using 2.766 bits per weight instead of 2.88.<br>DeepSeek V4 Flash 0731 quality at the model-default six experts per token ModelSizeBits per weightds4-eval-92 Lucebox ROCmFPX98.29 GB2.76682/92 Published reference—2.8882/92 The suite combines COMPSEC-17, AIME 2025, GPQA Diamond, and SuperGPQA. COMPSEC scored 17/17. All 92 evaluation items were excluded from calibration, and the outputs were rescored with the reference grader.<br>How the model fits on AMD Strix Halo<br>The GGUF is 98.29 decimal GB, or about 91.54 GiB. That is below the 96 GiB single-allocation window available to the Radeon 8060S in our 128 GB Strix Halo system, so the full model can stay resident on one local device. The 2.766 bpw figure is all-in: it divides the complete file, including its header and embedded codebooks, by the artifact&rsquo;s 284.33 billion parameters.<br>To reach that size, the mixed-precision layout described in PR #593 chooses each tensor&rsquo;s format from measured layer-output damage under real mixture-of-experts routing. The routed experts use learned per-expert codebooks: 2.5 bits per weight for gate/up experts and selected down projections, and 3.5 bits per weight for the remaining down projections. Attention, dense weights, and the shared expert stay at 4.25 bits per weight or higher.<br>WeightsFormatPrecision Routed gate/up experts + 15 down layersQ2_1_ROCMFP2_MIX2.50 bpw Remaining 28 routed down layersQ3_1_ROCMFP3_MIX3.50 bpw Attention, dense, and shared expert weightsQ4_0_ROCMFP4_FAST4.25 bpw Strix Halo configuration. The mixed formats require the expert weights and their codebooks to remain resident. These measurements use the Radeon 8060S iGPU and the 96 GiB allocation window on a 128 GB AMD Strix Halo system.<br>Six-expert performance on AMD Strix Halo<br>Every result in this article uses all six experts routed by the model for each token. In the published target-only measurement, the 98.29 GB target reaches 18.1 tok/s with fused decode disabled. This is also the routing configuration used for the 82/92 quality evaluation.<br>Target-only model-side decode on Ryzen AI MAX+ 395 with 128 GB unified memory ConfigurationRoutingDecodeCOMPSEC-17 Target onlyNative top-6, fused decode off18.1 tok/s17/17 The DSpark evaluation below keeps the same six-expert routing and enables fused decode. Fused decode changes how the work is executed; it does not reduce the number of routed experts.<br>DSpark speed evaluation: HumanEval, GSM8K, and MATH<br>On August 11, we ran a fixed 30-prompt evaluation: ten prompts each from HumanEval, GSM8K, and MATH. The Radeon 8060S handled both the 98.29 GB target and the draft on Strix Halo. The run used exact prefill, native six-expert routing, fused decode, DSpark q=4, temperature zero, and batch size one.<br>Fixed 30-prompt evaluation; server decode and acceptance are token-weighted within each row SuiteAudited scoreServer decodeDSpark acceptance HumanEval (10 prompts)10/1030.0 tok/s90.3% GSM8K (10 prompts)10/1026.4 tok/s75.4% MATH (10 prompts)9/1027.6 tok/s81.7% All 30 prompts29/3027.9 tok/s82.4% All 30 requests completed. Per-request draft acceptance ranged from 67% to 96%, with 82.4% token-weighted acceptance overall.<br>The 29/30 score is after auditing the saved responses without rerunning inference. The audit corrected one GSM8K answer-extraction error and one incorrect MATH gold label; one MATH response remained wrong. Both evaluator fixes have regression tests.<br>Run conditions....