DeepSeek V4 Flash: 284B model, up to 32 tok/s on AMD Ryzen AI MAX+ 395 | lucebox

July 2026<br>By Davide Ciffa<br>DeepSeek V4 Flash: 284B model, up to 32 tok/s on AMD Ryzen AI MAX+ 395<br>AMD-Powered Lucebox runs the full DeepSeek V4 Flash target locally on AMD Ryzen AI MAX+ 395 with 128 GB unified memory: up to 32.0 tok/s decode and roughly 250 tok/s with indexed sparse prefill. The code used for both paths is on Lucebox main.<br>284B<br>parameter target running locally on an AMD APU with 128 GB unified memory

up to 32<br>tok/s decode on the public LocalMaxxing entry

~250<br>tok/s indexed sparse prefill; 246.8–255.9 across the 8K validation cases

DeepSeek V4 Flash is a 284B mixture-of-experts model. Its target and draft both fit because the CPU and Radeon 8060S share the same 128 GB LPDDR5X pool. There is no discrete GPU, second machine, or remote inference service in this run: prompts and model execution stay on the local system.<br>The LocalMaxxing result<br>We submitted the run to LocalMaxxing. On July 18, its next-fastest DeepSeek V4 Flash entry for the Radeon 8060S was HipFire at 18.99 tok/s. The previous best in the site&rsquo;s Ryzen AI Max 395 unified-memory group was DwarfStar at 15.6 tok/s.<br>Published DeepSeek V4 Flash decode on Strix Halo · tok/s<br>DwarfStar Q2_KRyzen AI Max 395

15.60

HipFire MQ2 + MTPRadeon 8060S

18.99

Lucebox ROCmFPX + DSparkcurrent main

32.00

These are community submissions, not a controlled A/B test. Engines, quants, context sizes, and output lengths differ. Source: LocalMaxxing Ryzen AI Max 395 hardware page, checked July 18, 2026.

That puts our run 68.5% ahead of HipFire and at 2.05&times; the DwarfStar result. These are comparisons against the public LocalMaxxing entries shown above, not controlled A/B tests.<br>ROCmFPX: fitting 284B weights into 128 GB<br>ROCmFPX is not one quantization format. It is a family of block formats built around the AMD ROCm/HIP path. Each block holds 32 weights as packed low-bit codes plus one or two small scales. ROCmFP2 stores a block in 10 bytes, or 2.50 bits per weight; ROCmFP3 uses 3.50 bits per weight; and the fast ROCmFP4 layout uses 4.25.<br>For DeepSeek V4 Flash, we added the missing 2-bit format and its HIP kernels, then built a Strix-specific mixed-precision recipe. The enormous routed-expert gate and up matrices use ROCmFP2, expert down projections use ROCmFP3, and dense or more sensitive projections keep ROCmFP4 or higher precision. We used an importance matrix during quantization and kept the model&rsquo;s MTP head. The final 102.3 GB target works out to roughly 2.88 bits per parameter ; the filename says ROCmFP2 because that is the dominant format, not because every tensor is 2-bit.<br>On-disk model budget against 128 GB unified memory

102.3 GB ROCmFPX target(95.3 GiB)<br>11.3 GB DSpark draft(10.5 GiB)<br>14.4 GB remaining capacity before OS and runtime allocations

File sizes are shown in decimal GB to match the marketed 128 GB system capacity. Unified-memory residency is managed by ROCm and does not map one-for-one to this simple on-disk bar.

PieceMeasured configuration HardwareRyzen AI MAX+ 395, Radeon 8060S (gfx1151), 128 GB LPDDR5X TargetDeepSeek-V4-Flash-ROCMFP2-STRIX.gguf, 102.3 GB DraftDeepSeek-V4-Flash-DSpark-draft-Q4RMFP4-denseF16.gguf, 11.3 GB RuntimeROCm 7.2.4, HIP gfx1151, platform performance, Radeon high (2.9 GHz observed), q=4 verification cap Server context8,192 tokens in the published setup The format is also part of the speed result. At batch one, every generated token streams the active weights across all 43 layers, so decode is mostly limited by memory traffic. The ROCmFPX kernels read the packed blocks directly, expand their small codebooks in registers with AMD byte-permute instructions, and feed integer dot products without a separate codebook gather. In practice, the file layout and the GPU kernel are designed as one path.<br>Decode: up to 32 tok/s<br>ROCmFPX handles the weight traffic. We then added a DeepSeek-specific HIP decode path for the model&rsquo;s hyper-connections, attention, routing, and expert work. With no speculative draft, that target runs at 25.31 tok/s autoregressive.<br>DSpark is the next layer. With a q=4 batch, its small draft proposes up to three new tokens and the 284B target verifies four positions, including the current seed, in one fused pass.<br>01 · propose<br>DSpark draft<br>A compact three-layer draft proposes the next few tokens from captured target features.

02 · verify<br>q=4 target pass<br>The 284B target checks several positions together through the fused HIP graph.

03 · commit<br>accepted prefix<br>Correct proposals are committed in one step; the target repairs the first miss.

With a q=4 cap and adaptive width disabled, the public run reached 32.0 tok/s , 26.4% above the 25.31 tok/s autoregressive result. The gain varies with how many draft tokens the target accepts.<br>Same Lucebox target · decode tok/s<br>Autoregressive

25.31

DSpark q=4

32.00

Greedy decode, expert top-k 4, performance platform profile, high GPU performance...