Performance / measured on the production configuration

Over 2× the reported DGX Spark decode speed.

Lucebox generates 86 tokens per second on DeepSeek V4 Flash, the full 284B model, against 35.3 in the published single DGX Spark run. This page collects every number we publish for the workstation, the two engineering reports behind them, how to read them, and the AMD partnership the hardware comes from.

MoE · DeepSeek V4 Flash 86tok/s Token generation R9700 + Strix Halo · measured
MoE · DeepSeek V4 Flash 788tok/s @ 2K Prompt processing R9700 + Strix Halo · measured
Dense · Qwen3.8-27B 61.3tok/s · long prompt Token generation DFlash2 user run · prose 38.0 · count 217.9
Dense · Qwen3.8-27B 903tok/s @ ~6K Prompt processing DFlash2 user run · prose 38.0 · count 217.9
Built in partnership with AMD

Two models, three systems, every source linked.

Lucebox rows are measured on the workstation we ship. DGX Spark and Mac M5 Max rows are independent public runs with their own quantization and prompt lengths, linked on each row: point results, not a controlled comparison.

Dense

Qwen3.8-27B

One-user DFlash2 long-prompt decode · 61.3 tok/s on R9700

Qwen engineering ↗
01

Lucebox / R9700

DFlash2 user run · prose 38.0 · count 217.9

Source / One-user DFlash2 run

Prefill
903 tok/s @ ~6K
Decode
61.3tok/s · long prompt

Prompt processing

Point results · prompt lengths differ / tok/s

Higher is better

Token generation

R9700 uses one user's long-prompt result; public workloads differ / tok/s

Higher is better

MoE

DeepSeek V4 Flash

788 tok/s prefill · 86 tok/s decode

DeepSeek engineering ↗
01

Lucebox

R9700 + Strix Halo · measured

Source / Lucebox measurement

Prefill
788 tok/s @ 2K
Decode
86tok/s

Prompt processing

Reported 2K point results / tok/s

Higher is better

Token generation

Different test setups and quantization; separate reported results, not a controlled comparison.

Higher is better
An AMD-powered Lucebox combining a Radeon AI PRO R9700 with a Ryzen AI MAX+ 395

01MoE · 284B parameters · July 2026 report

DeepSeek V4 Flash: 3.63× the decode speed of one DGX Spark.

The full 284B model does not fit the R9700, so the engine gives the two AMD GPUs different jobs. The R9700 runs the dense path, the most frequently selected experts, the DSpark draft model and the working cache. Strix Halo holds the remaining experts in its 128 GB and runs the ones each token selects, in parallel, in the same process. We then put a Q2 build of the same model on one DGX Spark and measured both at 2K, 4K, 8K and 16K context.

51.1tok/s Token generation, serving median
3.63× One DGX Spark, 14.09 tok/s mean
415.52tok/s Prompt processing, DSpark loaded
ModelDeepSeek V4 Flash, 284B parameters, ROCmFPX compressed build (102.3 GB) with the 11.3 GB DSpark draft model
HardwareRadeon AI PRO R9700 (32 GB GDDR6) and Ryzen AI MAX+ 395 (128 GB unified memory), one process on ROCm 7.2.4
ProtocolOne request at a time, about 2K prompt tokens, 128 output tokens, temperature 0, median of three runs after two warmups
Matched sweep51.0 / 49.6 / 47.5 / 42.9 tok/s at 2K / 4K / 8K / 16K context against 14.18 / 14.24 / 14.04 / 13.89 on one DGX Spark, 3.39× on the mean
Range55.0 tok/s on a 53-token prompt with every draft token accepted, 45.5 tok/s median on a 510-token generation
Price$5,999 complete against $4,699 for one DGX Spark at publication: 28% more for about 2.8× the decode per dollar, and 36% less than two
Read the report
Qwen3.8-27B running on an AMD Radeon AI PRO R9700 with the DFlash2 block-diffusion drafter

02Dense · 27B parameters · August 2026 report

Qwen3.8-27B on one R9700: up to 227 tok/s, every token verified.

One Radeon AI PRO R9700 serves Qwen3.8-27B with the DFlash2 block-diffusion drafter from z-lab. The drafter proposes a whole block of tokens in one pass and the model verifies each one against its own logits, so acceptance changes speed and nothing else: blocks of 8, 12 and 16 produced byte-identical output on every prompt. Widening the block to 16 lifts the HumanEval average from 133.9 to 208 tok/s. The served file is Unsloth's UD-IQ4_XS exactly as downloaded, 8-bit class on every quality instrument we ran.

208tok/s HumanEval average, decode, block 16
227tok/s Fastest single request, on code
3.8× llama.cpp decode, same drafter and file
ModelQwen3.8-27B, Unsloth UD-IQ4_XS (13.27 GiB) as downloaded, z-lab DFlash2 drafter at block 16
HardwareOne Radeon AI PRO R9700, 32 GB, RDNA4, ROCm 7.2, no offload and no paging
ProtocolTen HumanEval-style prompts, 256 tokens, temperature 0, one request at a time, decode from each engine's own per-request timer
Against llama.cppSame card, same weights file, same drafter: 208.1 against 54.6 tok/s on decode (3.8×), 156.2 against 47.2 end to end (3.3×)
Long context128K context on the card; decode from 58 to 24 tok/s between 1.4K and 70K prompt tokens on prose, 142 to 45 tok/s on code
QualityMean KL divergence 0.018 against Q8_0, 94% top-1 agreement, HumanEval 151/164 and GSM8K 177/200, within noise of the 8-bit reference
Read the report

Built in partnership with AMD

Two AMD processors, one engine tuned for both.

Lucebox pairs the Ryzen AI MAX+ 395 (Strix Halo, 128 GB of unified memory) with the Radeon AI PRO R9700 (32 GB of GDDR6). AMD supplied the development hardware; we wrote the engine work that makes the pair fast: HIP kernels, RDNA4-tuned matmul tiles, the DFlash and PFlash speculative paths on Strix Halo, the DFlash2 drafter on the R9700, and the asymmetric expert parallelism that splits DeepSeek V4 Flash across both GPUs.

All of it is open source in the Lucebox engine, and every result on this page was measured on the same AMD silicon that ships in the workstation. AMD presented Lucebox at Advancing AI 2026 in San Francisco in July 2026.

AMD presented Lucebox at Advancing AI 2026
SiliconAMD Ryzen AI MAX+ 395 (Strix Halo) and AMD Radeon AI PRO R9700, both in every Lucebox
Memory128 GB LPDDR5X unified on the processor and 32 GB GDDR6 on the GPU, two pools the engine places the model across
StackROCm 7.2 and HIP kernels in the open source engine, on Ubuntu Server
Engine workDFlash and PFlash on Strix Halo, DFlash2 on RDNA4, DSpark drafting for DeepSeek V4 Flash, asymmetric expert parallelism across the two GPUs
HardwareDevelopment units supplied by AMD; production workstations use the same parts
EventAdvancing AI 2026, San Francisco, July 2026: AMD presented Lucebox

How to read the numbers

Point results, production configuration, reproducible.

01Point results, not a controlled comparison

Lucebox is measured on the production configuration. The DGX Spark and Mac M5 Max figures come from independent public runs with their own quantization, harness and prompt length, linked on each row. The reports add matched runs where we had the other hardware on the bench.

02Decode and prefill are separate measurements

Token generation (decode) and prompt processing (prefill) are timed separately and never added together. Speculative decoding pays in proportion to how predictable the next tokens are: code and math accept more draft tokens than prose, so the same workstation shows different decode speeds on different work.

03Every report keeps the figures it was published with

The engine ships new kernels every week. The tiles and the panel at the top carry the current measurement; each report keeps the numbers of its date, so the two can differ, and the newer figure is the one to expect.

04Reproducible on the workstation you receive

The engine is open source under Apache 2.0. Each report lists the exact model files, flags and commands, and the same build runs on every Lucebox: one request at a time, temperature 0, unless the report says otherwise.

Order

The configuration behind every number, from $5,999.

Both tiers ship the same R9700 and Ryzen AI MAX+ 395 with the engine and the qualified model profiles installed. The Business tier adds a year of Lucebox Engine Pro.

Order now See Engine Pro See components Read the docs Read the blog