AMD adds Day 0 support for MiniCPM-V 4.7 across GPUs and Ryzen AI

News

AMD says MiniCPM-V 4.7 was available on Radeon GPUs, Ryzen AI Max iGPUs and Ryzen AI NPUs at release, with validation and setup details.

AMD said MiniCPM-V 4.7 had Day 0 support across Radeon GPUs, Ryzen AI Max integrated graphics and Ryzen AI NPUs when ModelBest released the multimodal model on Oct. 14, 2026. The company framed the work as a deployment path for on-device and edge AI, while noting that its validation focused on functional correctness rather than speed or memory use.

Evidence and context

In an AMD developer article published Oct. 14, 2026, the company said MiniCPM-V 4.7 was supported on AMD Radeon GPUs, the integrated GPU in Ryzen AI Max processors and Ryzen AI NPUs based on the XDNA 2 architecture at the time of the model’s open-source release. The release comes shortly after OpenBMB introduced MiniCPM-V 4.7 on Oct. 12, 2026, according to the supplied background, and extends AMD’s earlier local-AI support work around the model family.

AMD said the model is offered in two variants, 1B and 35B-A3B, for image-and-text inference across AI PCs, automotive systems, robots, edge devices and servers. The article added that developers can choose the model size and inference path according to device resources and deployment conditions. AMD and ModelBest were described as working together on model, hardware and developer-resource integration.

The article’s validation section reported successful inference on three Radeon platform combinations: the Radeon AI PRO R9700 with RDNA 4 / gfx1201, the Radeon PRO W7900D with RDNA 3 / gfx1100, and the Radeon 8060S on Ryzen AI Max 395 with gfx1151. AMD said both the language model and visual projector used the ROCm backend and that outputs matched the test image. The company said the tests passed for both the 1B and 35B-A3B variants on the listed devices.

AMD also said the validation did not yet report throughput, time to first token or GPU memory usage. For the GPU workflow, the article described a llama.cpp HIP-backend build on Ubuntu or Debian, using GGUF model files and a matching visual projector file, followed by command-line inference or launch of an OpenAI-compatible local server.

On the NPU side, AMD said FastFlowLM v1.0.8 adds support for MiniCPM-V-4.7-1B inference on Ryzen AI NPUs. Supported platforms include Ryzen AI PCs with XDNA 2 NPUs, covering AMD Ryzen AI Max, Ryzen AI 400 and Ryzen AI 300 series. AMD said the NPU model tag is minicpm-v-4.7:1b, uses Q4_K quantization in NPU2 format and includes model.q4nx and vision_weight.q4nx files totaling about 2.2 GB.

AMD’s NPU example used an AMD Ryzen AI Max+ 395 system running Windows 11 with FastFlowLM v1.0.8 and NPU driver version 32.0.20102.3930. In that validation, AMD said the model handled both text conversation and image understanding, and reported an approximate time to first token of 1.2 seconds for a single image, with decoding speed ranging from about 44 tokens per second to 23 tokens per second at context lengths from 1K to 32K. Those figures were supplied by AMD and apply only to the test setup described in the article.

The article closes by directing developers to AMD’s developer program and model links, but it does not provide a separate public catalyst date beyond the Oct. 14 publication of the support article. No independent benchmark data were supplied in the source material.