Tiny's $30K tinybox swaps RAM for 16 SSDs to run a 2.8T model

What exactly is the tinybox Red Kimi?

It's a desktop workstation from Tiny Corp built around four AMD Radeon AI Pro R9700 GPUs and sixteen 1 TB NVMe SSDs, designed to run a 2.8-trillion-parameter model locally.

Tiny Corp has never been shy about building odd machines, but the tinybox Red Kimi is its most ambitious yet. The $30,000 box pairs four AMD Radeon AI Pro R9700 graphics cards — each with 32 GB of GDDR6 at 640 GB/s — for a total of 128 GB of GPU memory and 2,560 GB/s of aggregate bandwidth. A 32-core AMD Epyc Genoa processor and 32 GB of DDR5 system memory round out the package.

The real party trick is storage. Sixteen 1 TB NVMe SSDs form a RAID array rated at 238.4 GB/s. Tiny Corp says its tinygrad software now includes a driver that lets the AMD GPUs read those drives directly, bypassing the CPU. The company pitched the result on X as "frontier models at home at usable speeds for a good price."

At 56 pounds and measuring 12 inches wide, 19.25 inches tall and 18 inches deep, the Red Kimi is not exactly a desktop replacement. Tiny also warns that the R9700's blower fans are "audible," pushing noise from under 50 dB to under 60 dB.

Tinybox Red Kimi: specs, price and availability

Why swap DRAM for SSDs?

Because Kimi K3 is enormous — roughly 1.56 TB in its native checkpoint — and filling that much DRAM would cost far more than flash.

Moonshot's Kimi K3 is a mixture-of-experts model with 896 routed experts, of which 16 are active per token, plus two shared experts. That works out to 104 billion active parameters per token. At MXFP4's effective 4.25 bits per weight, each token requires reading at least 55 GB of weights.

The four GPUs can hold about 8% of the model's 1.56 TB checkpoint. The rest has to live somewhere, and Tiny Corp chose considerably cheaper flash over DRAM. The concept isn't entirely new: researchers from Nvidia and the University of Illinois presented GPU-initiated SSD access in the BaM system architecture at ASPLOS 2023, treating storage as another tier of accelerator memory. Tiny's offering is a working version on AMD workstation cards, inside a box you can actually order.

The trade-off is speed. If every byte of those 55 GB came from the SSD array, the system would manage only about four tokens per second. To hit 40 tokens per second, the GPUs would need to move roughly 2.2 TB/s of weights — about 86% of their combined memory bandwidth — while the SSDs contribute at most 6 GB per token, or around 11% of what each token needs.

The 40 tokens per second claim: plausible or wishful?

It's mathematically possible, but only if the GPUs' 128 GB already contains nearly everything each token touches — and Tiny hasn't published the details.

Tiny Corp claims the Red Kimi can run Kimi K3 at 40 tokens per second. Nobody outside the company has tested that figure. The math suggests it could work if expert selection is predictable enough to cache well, or if Tiny is running a lower-precision quantization than Moonshot's native format, speculative decoding, or a mix of the three.

But Tiny hasn't published the precision, context length, batch size, or method behind its number. The aggregate bandwidth figure also assumes the workload is split evenly across four cards that communicate over PCIe. Until an independent benchmark appears, treat 40 tokens per second as a manufacturer claim, not a result.

How does it compare to Apple's Mac Studio and Nvidia's DGX Station?

The Red Kimi is the cheapest way to run Kimi K3 at native precision on a single desk, but it's only $201 less than a maxed-out Mac Studio.

Apple's Mac Studio with M5 Ultra and 512 GB of unified memory starts at $22,299 and reaches $29,799 with a 16 TB SSD, according to Macworld. That's just $201 shy of Tiny's box. The Mac offers four times the fast memory at 1.2 TB/s, but 512 GB still can't hold K3's native checkpoint. As we reported, Apple's 512GB Mac Studio costs $29,799, with the memory upgrade alone at $15,500.

Nvidia's GB300-based DGX Station went on sale for almost $91,100 with 748 GB of memory, which also falls short. AMD's Threadripper Halo Station, with up to 2.6 TB of combined memory, would fit the model comfortably, but it isn't due until 2027, and industry estimates put a full system at around $150,000.

Against such competition, the Red Kimi is the cheapest way to maintain Kimi K3's native precision on a single desk. The catch is that it gets there by asking flash to do the job of memory, and every token per second it delivers depends on how well that bet pays off.

Who is this actually for?

Enthusiasts and small labs who want to run massive open models locally without renting cloud GPUs.

The Red Kimi is not for everyone. At $30,000, it costs as much as a decent car. But for researchers, startups, or wealthy hobbyists who need to run Kimi K3 or similar models without sending data to a cloud provider, it's the only single-desk option that maintains native precision.

Tiny Corp has a track record of shipping niche hardware to a devoted community. The Red Kimi continues that tradition: a machine that makes trade-offs no mainstream vendor would dare, in exchange for capabilities that are otherwise out of reach. Whether the 40 tokens per second materializes in real-world use is the question that will decide its fate.

Availability and pricing

The tinybox Red Kimi is available to order now directly from Tiny Corp's website, with a starting price of $30,000. Shipping availability for the US, UK and EU has not been detailed, but Tiny Corp has historically shipped its tinybox systems internationally.

At $30,000, the Red Kimi sits in the same price bracket as Apple's maxed-out Mac Studio, but offers a very different set of trade-offs. Potential buyers should factor in power, cooling and noise requirements before committing.

Tiny Corp's tinybox Red Kimi is a bold experiment in local AI hardware, trading raw memory bandwidth for massive SSD storage. It's the cheapest way to run Kimi K3 at native precision on a desk, but the performance claims remain unproven. For now, it's a machine for early adopters and enthusiasts who can afford the gamble.