Verified reference stack

Research — Private Research Assistant — Apple 32 GB

Gemma 3 12B int4 weights reported by Google at ~6.6 GB; 32 GB unified memory offers substantial headroom for a bounded-context local assistant.

Application

Khoj

Open tool profile →
Runtime

Ollama

Open runtime profile →
Model

Gemma 3 12B IT (QAT/int4 profile)

Hardware

Apple Silicon 32 GB unified memory

GOOD
Privacy

Fully local configuration

License / terms

Gemma Terms of Use — professional/legal-entity use allowed subject to terms and prohibited-use policy

Evidence

Continue the decision

Put this reference stack in context

A reference configuration is most useful when you compare it with alternatives and then apply your own hard constraints.

Freshness rule: Material changes set this stack to REVALIDATION_REQUIRED.