A 284-billion-parameter AI model now runs on a single laptop. Here is what DeepSeek V4 Flash is and why engineers are talking about it.

V4 Flash is a mixture-of-experts model. Only 13 billion of its 284 billion parameters are active per token, so it behaves more like a small model during each step while drawing on a huge knowledge base.

It is MIT-licensed, which means anyone can use, modify and deploy it commercially. That openness is rare for a model of this scale.

The buzz comes from a new engine called DwarfStar 4, built by Redis creator Salvatore Sanfilippo. It squeezes V4 Flash into about 76GB using asymmetric quantization: expert layers drop to 2-bit while routing layers stay precise.

On a MacBook Pro M3 Max it hits about 26.7 tokens per second, fast enough for real work. The engine also ships a coding agent, tool calling and a disk-persisted memory that survives between sessions.

Why it matters: frontier AI is leaving the data centre. A model this capable, running privately on hardware you own with no API bill, changes who can build with AI and where.