Radar ·
Local LLM benchmarks on the Ryzen AI Max+ PRO 495 and a return decision
A reviewer reported 18 to 58 tokens per second with four large local models on llama.cpp with Vulkan, then returned the hardware because of its cost.
According to the publication dated 18 October 2026, a reviewer ran four large language models on a Minisforum Ryzen AI Max+ PRO 495 with 192 GB of memory, using llama.cpp with a Vulkan backend. The text reports throughput of 18 to 58 tokens per second and shows the memory use of large local models on unified-memory APUs. After the tests, the reviewer returned the hardware because of its cost.
The relevance to Rota Nacional is indirect: the platform is a hosted model access API, compatible with OpenAI and Anthropic, and not software for running models on the customer's own hardware. When using AI to study this material, do not paste internal configurations, invoices or machine credentials into prompts. The personal data detector covers CPF, CNPJ, e-mails, phone numbers and names, but it does not replace reviewing internal data. To check the original, find the article by its title and confirm the date, hardware, llama.cpp version and described methodology.
Original source: Benchmarks de LLM local no Ryzen AI Max+ PRO 495 e decisão de devolução