Lucebox partners with AMD on a $6,499 local AI inference box

Why it matters

Lucebox is testing whether software tuned for mismatched consumer processors can give AMD a credible position against Nvidia's integrated desktop AI systems.

Lucebox founder Alessandro "Sandro" Puppo (@pupposandro) announced a partnership with AMD on July 30th to sell a local AI inference computer that combines a Radeon AI PRO R9700 discrete GPU with AMD's Ryzen AI MAX+ 395, commonly known as Strix Halo.

Puppo, who previously worked at Notion and helped build Cua, a Y Combinator-backed infrastructure project for computer-use agents, is taking the same underlying bet into hardware: AI agents will need a private, predictable source of inference that developers can run continuously without sending every request to a cloud API.

The San Francisco-based Lucebox lists the AMD-powered machine at $6,499 through August 31st, after which the price rises to $7,900. The system includes 128GB of LPDDR5X unified memory attached to the Strix Halo processor, 32GB of GDDR6 memory on the R9700, a 2TB NVMe drive and an aluminum 11.97-liter enclosure. Ubuntu, tuned models and Lucebox's inference engine arrive preinstalled, according to Lucebox.

The product exposes OpenAI- and Anthropic-compatible endpoints, allowing developer tools and agent frameworks to send requests to the machine using interfaces modeled on cloud APIs. Lucebox says the system can operate offline once model weights are loaded.

Lucebox splits one model across two different AMD processors

The technical centerpiece is Lucebox's approach to heterogeneous computing. Instead of dividing each model layer evenly between two identical accelerators, Lucebox assigns different work to the R9700 and Strix Halo based on their memory capacity and speed.

In a technical report published alongside Puppo's announcement, Lucebox said it loaded a compressed, 102.3GB build of the 284-billion-parameter DeepSeek V4 Flash model across the two processors. The R9700 stored the dense path, frequently selected expert blocks, an 11.3GB speculative-decoding helper model and the working cache. Strix Halo's larger unified memory held the remaining experts.

Both processors execute selected experts concurrently before the R9700 combines their outputs. The architecture is designed around mixture-of-experts models, which activate only a fraction of their total parameters for each token. Lucebox calls the technique asymmetric expert parallelism.

Lucebox reported a median generation rate of 51.1 tokens per second for a single request using an approximately 2,000-token prompt and 128 generated tokens. The result was the median of three measured requests after two warmup runs. Puppo said on X that the result was 3.63x the performance Lucebox measured on one NVIDIA DGX Spark.

That headline ratio needs qualification. Lucebox measured the DGX Spark at an average 14.09 tokens per second across four context lengths, while the 51.1-token Lucebox figure came from a separate serving test. In Lucebox's matched sweep at 2,000, 4,000, 8,000 and 16,000 tokens of context, its machine averaged 47.75 tokens per second, producing a 3.39x speedup over the DGX Spark result.

The two systems also ran different model compression formats and software configurations optimized for their respective hardware. Lucebox used a ROCmFPX build, four routed experts per token and speculative decoding. The DGX Spark used a Q2-compressed model. Lucebox describes the result as a whole-system comparison rather than a GPU-only test.

The benchmark was run by Lucebox rather than an independent testing lab. Lucebox has published the test setup, model names and implementation through its open-source repository, giving developers a path to inspect and reproduce the work.

AMD gets another route onto AI developers' desks

The partnership gives AMD a small but focused hardware vendor building directly around the combination of a high-memory APU and a discrete Radeon GPU. AMD positions the R9700, which has 32GB of video memory and uses its RDNA 4 architecture, for local inference and memory-intensive development workloads.

Lucebox's approach also addresses a persistent disadvantage for AMD in developer AI systems: much of the inference software market has been built around Nvidia's CUDA stack. Lucebox is attempting to close part of that gap through model-specific kernels, speculative decoding and placement logic written for the exact AMD hardware configuration.

The pricing puts Lucebox between one and two DGX Spark units. Nvidia currently lists one DGX Spark at $4,699 with 128GB of unified system memory and 4TB of storage. At their listed prices, the Lucebox costs about 38% more than one DGX Spark and about 31% less than two.

Puppo's wager is that developers running sustained agent workloads will pay the premium for higher measured throughput, local data control and a preconfigured software stack. The hardware is only half of that proposition. Lucebox's ability to keep tuning individual open models across AMD's two processors will determine whether the box remains faster after the benchmark model changes.