AMD showed how to launch the thinking model Deepseek R1 on a regular PC

AMD has published instructions for the local launch of the advanced Deepseek R1 AI model with the ability to reflect on user PCs. True, a system is needed on the Ryzen processor with AI accelerator (NPU) XDNA, or a computer with a Radeon RX 7000 desktop video card. Also, the Adrenalin driver 25.1.1 is required to start.

Image Source: AMD

It is claimed that only recently a “highly effective” small version of the Deepseek R1 AI model was presented, which is compact enough to work on consumer equipment. It is worth noting that in general, the deepseek models are characterized by small queries to the iron-for example, the DeepSeek-V3 model was originally studied on a cluster of only 2048 NVIDIA H800 accelerators.

The AMD instructions describe everything that is needed for the local launch of Deepseek R1 on the user system based on AMD hardware. Lm Studio has a installer of large language models in one click specially designed for Ryzen AI processors. AMD users can use it for installation R1. In the video, the company also shows how the application should be configured for specific equipment, and also explains how the maximum number of LLM parameters is supported by one or another equipment.

The latter largely depends on the amount of accessible memory. For example, the Radeon RX 7600 XT, RX 7700 XT, RX 7800 XT, RX 7900 GRE and RX 7900 XT support models to Deepseek-R1-Distill-QWEN-14B with 14 billion parameters. The flagship map Radeon RX 7900 XTX allows you to launch models to Deepseek-R1-Distill-QWEN-32B with 32 billion parameters. And the Radeon RX 7600 model, which has only 8 GB of memory on board, supports the launch of models to Deepseek-R1-Distill-Llama-8B with 8 billion parameters.

Laptops based on the processors Ryzen 8040 and Ryzen 7040 with 32 GB of RAM, as well as systems based on the Ryzen AI 9 HX 370 and Ryzen Ai 9 365 processors, equipped with RAM, can run models to Deepseek-R1-Distill-Lalama -14b. Laptops based on Ryzen Ai Max+ 395 support models to Deepseek-R1-Distill-Llama-70B with 70 billion parameters, but only if there are 64 or 128 GB of RAMs. Versions of laptopes with 32 GB of memory can launch models to Deepseek-R1-Distill-QWEN-32B with 32 billion parameters.

The new model of artificial intelligence Deepseek R1 defeated the world overnight, since the costs of its training were 11 times lower than that of the advanced AI models of competitors. Two days ago, it caused a record loss of market capitalization of NVIDIA in the amount of $ 589 billion. The DeepSeek R1 model relies on extreme optimization levels to ensure 11 times an increase in efficiency. DEPSEEK R1 works not only on NVIDIA and AMD accelerators, but also reports supporting Huawei Ascent accelerators.