# DFRobot DX-M1 brings up to 25 TOPS of AI inference to an M.2 module > DFRobot's $140 DX-M1 is designed for local computer vision and other inference tasks, with a claimed 25 TOPS at INT8 and power consumption of 2 to 5 watts. It uses an M.2 slot but does not replace an SSD. > > OpenCheese AI · 2026.11.09 · Web page: https://opencheese.dev/blog/dfrobot-dx-m1-brings-up-to-25-tops-of-ai-inference-to-an-m-2-module DFRobot's DX-M1 offers local AI processing without a powerful graphics card or a cloud connection, but it asks for something compact computers may have little of: an M.2 slot. The $140 accelerator promises up to 25 TOPS at INT8 precision while consuming 2 to 5 watts, with computer vision and robotics among its intended uses. ## A small module for inference The DX-M1 connects over PCIe 3.0 and is built to run trained neural network models rather than train them. DFRobot identifies object recognition, video stream analysis, robot navigation and automated visual inspection as suitable tasks. The advertised 25 TOPS is an INT8 performance figure. It is a useful specification for understanding the device's intended workload, though the report does not provide results for particular models or real-world applications. The module has 4 GB of LPDDR5 memory for neural network weights and 128 GB of QSPI NAND flash for firmware and local storage of compiled models. Those resources serve the accelerator. They do not turn it into a general-purpose replacement for a computer's main storage. ## Software support and model limits DFRobot's DXNN SDK supports PyTorch, TensorFlow, TensorFlow Lite and ONNX. That gives developers a route to use models from those frameworks with the module, although the report does not detail the conversion process or model compatibility. The DX-M1 is not intended to run large language models directly. Its 4 GB of memory is a constraint for heavier LLMs, and its stated role is inference with trained models, particularly for computer vision. ## The M.2 trade-off A buyer needs to account for the slot as well as the power draw. The DX-M1 occupies an M.2 position that could otherwise hold an NVMe drive, and its onboard flash is not intended to store the operating system or replace an SSD. That makes the module a more natural fit for compact computers, industrial systems and robots that need local, low-power vision processing and have room for both acceleration and storage. The accelerator may be small, but the slot still has a day job. ## Related posts - 2026.11.01 · [Asus adds a 64 GB Ascent GX10 for local AI, with no official price yet](https://opencheese.dev/blog/asus-adds-64-gb-ascent-gx10-for-local-ai-no-official-price-yet.md): Asus has added a 64 GB configuration to its Ascent GX10 mini computer, built on Nvidia's DGX Spark platform. Pre-orders are open in Germany, Austria and Switzerland, but official pricing and shipping dates are still missing. - 2026.11.10 · [Jensen Huang withdraws from ElevenLabs Summit as Nvidia investigates relationship](https://opencheese.dev/blog/jensen-huang-withdraws-from-elevenlabs-summit-as-nvidia-investigates-relationshi.md): Nvidia CEO Jensen Huang has dropped out of the ElevenLabs Summit and other events after admitting an affair with Megaspeed executive Alice Huang. Nvidia is investigating the relationship amid allegations that she worked for Chinese intelligence. - 2026.11.09 · [Apple Plans to Turn Siri Into an AI Agent for Third-Party Apps](https://opencheese.dev/blog/apple-plans-to-turn-siri-into-an-ai-agent-for-third-party-apps.md): Apple is preparing a more autonomous Siri that could book tables, order food and carry out multistep tasks in its own apps. Some features could arrive in December 2026, with third-party integrations expected in spring 2027.