Minisforum says a single unit can handle models in the 300-billion-parameter class. In its own tests, one machine ran DeepSeek V4 Flash, a 284-billion-parameter model, compressed to 4-bit precision (a common trick called quantization that shrinks a model to fit in less memory), at about 12 tokens per second. A token is roughly a word or part of a word, so that’s a readable chat speed.
Link two units together and Minisforum says they ran Qwen3.5-397B at around 16 tokens per second. A cluster of four loaded the full 380GB version of DeepSeek-R1 671B, though the company didn’t give a speed for that test.
Those speeds are possible because both models use a mixture-of-experts design. Instead of using every part of the model for each word, they switch on only the specialist sections they need. DeepSeek V4 Flash uses about 13 billion of its 284 billion parameters per token, and Qwen3.5-397B uses about 17 billion. Memory size decides whether a model fits at all, while memory speed and those active parameters decide how fast it answers. With roughly 273 GB/s of memory bandwidth, the box can still manage double-digit speeds.
Keep in mind that these numbers come from Minisforum, not independent testing. The company also pitches the machine at creative professionals for jobs like 3D rendering and AI-assisted design, not just chatbots.