Inkling 2 Mini runs a local Mac assistant: 26.4B params, only 3.4B active

simonguozirui · x · 2026-11-21

Developers with early access to Thinking Machines' Inkling 2 Mini built a fully on-device personal AI assistant: the model was served via Ollama on a single Mac mini and connected to OpenJarvisAI to schedule meetings, plan trips, draft texts, send emails and manage an inbox — all locally.

Architecturally notable: 26.4B total parameters with only 3.4B active per token; the middle 12 layers run three times with the same weights, yielding 5.4B weight activations per token. The model writes code, interprets images, and handles multistep requests, showing small MoE models can power on-device agents.

Original post →

More from Infra

Infra channel →