이미지: METAL LAB 생성
Summary
- Tiny corp released a demo showing an AMD 7900XTX connected via USB3 running Qwen 3.6 27B at 34 tokens per second.
- The eGPU dock that supports this method ships on the 12th, with fully open-source firmware, the company said.
- It added that it included an extra USB port for serial communication so the device can be recovered even if the firmware becomes corrupted.
- 구동 모델
- Qwen 3.6 27B
- 가속기
- AMD Radeon 7900XTX
- 연결 방식
- USB3
- 측정 속도
- 초당 34토큰
- 도크 출시
- 12일, 펌웨어 100% 오픈소스
- 복구 기능
- 시리얼용 USB 포트 추가로 브릭 방지
- 공개 주체·시점
- the tiny corp, 2026년 8월 10일 X 게시
34 tokens per second. That's fast enough for text to stream out faster than a person can read. But the setup that produced this speed is unusual. The computer and graphics card running a 27-billion-parameter-class model weren't connected through a PCIe slot — just a plain USB cable.
What was shown
On August 10, tiny corp posted a video on X showing an AMD 7900XTX connected via USB3 running Qwen 3.6 27B. The company wrote that this works on "any computer made in the last decade." It added that the external GPU (eGPU) dock handling this connection ships on the 12th, with fully open-source firmware, and that it included an extra USB port for serial communication so the device won't die even if a firmware update goes wrong.
Why USB3 is surprising
Connecting an external graphics card to a laptop has traditionally been tricky. Graphics cards are designed to plug into a motherboard's PCIe slot and exchange tens of gigabytes per second, and extending that connection outside the case required a standard like Thunderbolt or OCuLink that carries the PCIe signal directly. Thunderbolt ports have only appeared on mid-to-high-end laptops in recent years. USB3, by contrast, has been present on virtually every PC since the early 2010s.
The bandwidth gap is still large. But language model inference doesn't demand bandwidth the way gaming does. Once model weights are loaded into GPU memory, what flows back and forth afterward is mostly input and output tokens. The heavy computation all happens between the VRAM and compute units inside the card itself. The 7900XTX is a high-end gaming card with 24GB of VRAM, and a 27B-class model fits within that when quantized to around 4 bits. In other words, even if the "pathway for moving cargo" is narrow, performance doesn't drop much once the cargo has been moved.
The company behind it
Tiny corp was founded by George Hotz, known for iPhone jailbreaking and open-source autonomous driving projects. The team built the deep learning framework tinygrad and has sold tinybox, a complete AI workstation built on top of it. This team has consistently focused on rebuilding software stacks from scratch outside the NVIDIA CUDA ecosystem, particularly for AMD cards. They're known for building their own implementations that bypass manufacturer drivers to run AMD GPUs. Open-sourcing the dock's firmware entirely and adding a recovery serial port is an extension of that same philosophy — reducing black boxes in hardware so users can inspect failures themselves.
So what changes
Over the past year or two, a flood of open models in the 20-30B range has emerged — a size that fits on a single high-end gaming graphics card. The problem was where to plug that card in. For people without a desktop tower, or who only use a work laptop, local inference has always remained one step out of reach. The connection standard now brought down to USB3 lowers that barrier. It creates the option of turning an old mini PC or aging laptop, without discarding it, into a personal inference machine just by adding a card and a dock.
That said, what's confirmed so far is a single demo the company released and the figure of 34 tokens per second. Initial loading time to bring a model into memory, behavior with longer contexts, and scalability when multiple cards are attached are all things that will only become clear once the actual product ships on the 12th and users try it in their own setups.
