Science & Tech · AI · 4 hrs ago
Tiny Corp's $30,000 workstation aims to run Kimi K3 at home
Tiny Corp has introduced a $30,000 workstation designed to run Moonshot AI's Kimi K3 model on a desk.
The system combines four AMD GPUs with 128 GB of memory and sixteen 1 TB SSDs.
Kimi K3 is a very large model, with weights measuring about 1.56 TB, far more than the GPUs can hold at once.
Tiny Corp says its software lets the GPUs access the SSD array directly, which is rated to read data at 238.4 GB per second.
The company claims the machine can generate 40 tokens per second, but no one outside the company has tested that speed.
Reaching it appears to depend on keeping nearly all the weights needed for each token in GPU memory, so the SSDs cannot supply most of the data on their own.
The workstation is available to order, though it weighs about 56 pounds and its GPU fans can be audible.
Tiny Corp’s $30,000 tinybox Red Kimi is designed to run Moonshot AI’s 2.8-trillion-parameter Kimi K3 model on a workstation.
It combines four AMD Radeon AI Pro R9700 GPUs, with 128 GB of total GPU memory, and sixteen 1 TB NVMe SSDs.
Tiny Corp says the system can generate 40 tokens per second, but TechRadar says nobody outside the company has tested that figure.
TechRadar’s analysis says reaching that speed would require the GPUs to cache nearly all the weights used for each token in their memory.
The system uses Tiny Corp’s tinygrad software to let the GPUs access the SSD array directly.
- Who
- Tiny Corp is offering the tinybox Red Kimi for users seeking to run Moonshot AI’s Kimi K3 model.
- What
- It is a $30,000 workstation with four GPUs and sixteen SSDs. Tiny Corp claims it can run Kimi K3 at 40 tokens per second.
- When
- TechRadar published its report on November 5, 2026.
- Where
- The workstation is designed to run at home or on a desk rather than in a data center.
- Why
- It is intended to run a 2.8-trillion-parameter model locally, using SSD storage to supplement the GPUs’ memory.
Tiny Corp
TechRadar
Performance claim
Tiny Corp
Tiny Corp claims the workstation can run Kimi K3 at 40 tokens per second.
TechRadar
TechRadar says the claim has not been independently tested and would depend on caching nearly all the weights used for each token in GPU memory.
Storage approach
Tiny Corp
Tiny Corp’s system uses GPU-directed access to an SSD array to support running the model.
TechRadar
TechRadar calculates that reading all required weights from the SSDs would support only about four tokens per second.
Researchers from Nvidia and the University of Illinois presented GPU-initiated SSD access in the BaM system architecture.
TechRadar reported on Tiny Corp’s tinybox Red Kimi and its claimed performance.
- Price
- $30,000
- GPUs
- Four AMD Radeon AI Pro R9700 cards
- GPU memory
- 128 GB GDDR6
- Storage
- Sixteen 1 TB NVMe SSDs
- SSD array speed
- 238.4 GB/s
- Claimed model speed
- 40 tokens per second
What people are saying
frontier models at home at usable speeds for a good price