Tiny Corp unveils new AI workstation with monstrous 16 TB SSD array and 128 GB GPU memory

Tiny Corp has introduced a new configuration of its Tinybox AI workstation for running large language models locally. Dubbed the tinybox Red Kimi, the $30,000 machine features four workstation-grade GPUs, a massive high-speed SSD array and enough computing power to reportedly run Moonshot AI’s Kimi K3 40 tokens per second.

The new configuration has appeared on Tiny Corp’s official website. At its heart is a 32-core AMD Epyc Genoa processor (exact SKU not specified), paired with 32 GB of DDR5 system memory and four Radeon AI Pro R9700 GPUs. Each graphics card contributes 32 GB of GDDR6 memory and 640 GB/s of bandwidth, bringing the combined figures to 128 GB and 2,560 GB/s, respectively.

However, the most interesting part of the Red Kimi is its gargantuan storage array. Tiny Corp has equipped the workstation with 16 individual 1 TB NVMe SSDs, arranged in a high-speed RAID configuration. The company advertises 238.4 GB/s of storage bandwidth, substantially higher than what conventional desktop storage configurations can deliver. The company has also included an infographic to show how the storage works. 

Tiny Corp also claims to have developed a custom driver that allows AMD GPUs to directly control the NVMe drives. This matters for running enormous AI models such as Kimi K3, which has 2.8 trillion parameters and cannot fit entirely in the workstation’s 128 GB of GPU memory. The high-bandwidth storage arrangement is intended to help overcome that limitation by making model data accessible from SSDs.

We previously covered Tiny Corp’s revamped Tinybox workstation earlier this month. That system introduced a configurable G4/G5 platform, with options ranging from four Radeon RX 9070 XT GPUs to Nvidia’s professional RTX cards. It also supported up to 16 TB of additional RAID storage, albeit through a more conventional four-drive configuration.

The Red Kimi is listed for $30,000 on Tiny Corp’s configurator. Like other Tinybox machines, it is built to order, with estimated shipping times of one to eight weeks. Orders are paid for by bank transfer. Tiny Corp also offers a cheaper $27,000 red flash configuration with six Radeon AI Pro R9700 GPUs for 192 GB of combined VRAM, but without the Red Kimi’s specialised storage setup.

It is worth noting that the advertised 40 tokens/s Kimi K3 performance remains a manufacturer claim, with no independent end-to-end benchmarks available yet. Nevertheless, the unusual combination of high-speed storage and AMD hardware could make the Red Kimi an interesting alternative to something like the 512 GB RAM variant of the Mac Studio