GEEKOM is showing how far a Mini PC setup can be pushed for local AI workloads, with four A9 Mega systems working together to run DeepSeek V4 Flash. Rather than putting the workload on a traditional data-center server, GEEKOM has connected four compact machines through USB4 to create a distributed AI platform that can be deployed on a desk or in a small business environment.
Each GEEKOM A9 Mega is powered by AMD’s Ryzen AI Max+ 395, a processor with 16 Zen 5 CPU cores and Radeon 8060S graphics. The chip also uses unified memory, allowing CPU and GPU resources to work with the same memory pool. Combined across four systems, the configuration gives GEEKOM a compact way to handle larger local AI workloads without building a conventional server system.
The DeepSeek V4 Flash deployment uses Ubuntu, AMD ROCm, and DwarfStar to distribute the optimized model across the four machines. GEEKOM says the systems communicate through USB4, avoiding the requirement for a proprietary high-speed networking switch. An OpenAI-compatible API is also available, allowing applications and AI agents to interact with the cluster without requiring a completely different software interface.
The setup is particularly aimed at organizations that want to keep AI workloads inside their own infrastructure. Instead of sending prompts, documents, source code, credentials, or intermediate results to a public cloud, companies can process this information locally. That could make the configuration useful for internal knowledge assistants, document analysis, source-code review, local RAG workloads, research, and workflow automation.
GEEKOM also says the platform can support agent-based workloads such as Hermes Agent. These workloads can combine tools, memory, policies, logs, code, and retrieved information before an AI agent takes action. The company reports that the configuration has operated with context sizes reaching 250K tokens, which could help with large documents and sizeable software projects.
In testing, the four-system configuration achieved around 14.61 tokens per second at single concurrency, while P95 time to first token was approximately 0.42 seconds in the 32- and 128-token tests. GEEKOM says the value of the cluster is not limited to raw text-generation speed, with the additional systems providing more capacity for long prompts and demanding inference workloads.
The modular approach also means users do not necessarily have to deploy four systems from the beginning. A single A9 Mega can operate on its own, while additional units can be added for distributed inference as workloads increase. This gives businesses a way to expand local AI capacity without replacing their existing hardware setup.
- Four GEEKOM A9 Mega Mini PCs run DeepSeek V4 Flash together.
- Each system uses an AMD Ryzen AI Max+ 395 processor.
- Ryzen AI Max+ 395 includes 16 Zen 5 CPU cores and Radeon 8060S graphics.
- USB4 connects the Mini PCs into a distributed AI platform.
- Ubuntu, AMD ROCm, and DwarfStar manage the deployment.
- The setup supports an OpenAI-compatible API.
- GEEKOM reports context sizes of up to 250K tokens.
- Reported performance reaches about 14.61 tokens per second at single concurrency.
- The cluster can be expanded from one system to four.
- Target workloads include private AI assistants, RAG, coding, document analysis, and research.
For organizations looking for a compact alternative to a conventional AI server, GEEKOM’s approach puts four Ryzen AI Max+ 395 Mini PCs into a distributed configuration designed around local inference. The combination of DeepSeek V4 Flash, ROCm, USB4, and DwarfStar could make this type of setup an interesting option for offices, laboratories, classrooms, and edge computing environments.
