Chip off the old bot? Post says agent-led accelerator runs Qwen

An open-source project called openTPU is putting AI agents to work on the hardware side of AI, according to a Japanese-language post describing its accelerator design. The post says the project ran Qwen models with actual weights on an FPGA, a programmable chip, rather than merely testing a design in simulation.

The writer relays the project developer’s measurement of about 21.6 tokens a second for Qwen3-0.6B using int8 weights. Running a model on hardware is the notable step here: the project is testing an accelerator intended for AI inference, not announcing a new Qwen model.

There are limits to the bot-built-chip story. The post says it found no third-party reproduction of the speed or compatibility results, and that the project documentation does not spell out how much work humans did. It also cautions that a working FPGA demonstration is not the same as a custom chip ready for production.