The Largest Local Models: 4B to 9B on Your Phone

The free edition runs models up to the mid tier — roughly 3B parameters. The unlock opens the high and flagship tiers, which is where local models stop being impressive-for-a-phone and start being genuinely useful.

What that means in practice

The high tier is 4B-class: Qwen3.5 4B, Gemma 3 4B (quantisation-aware trained, so it holds up far better than a naive 4-bit quantisation), Phi-4 Mini, and Gemma 3n E2B. It wants around 8 GB of RAM. The flagship tier is Qwen3 8B and Qwen3.5 9B, which needs 12 GB or more.

The difference you feel is not eloquence — small models are already fluent. It is following instructions. A 4B model keeps track of a multi-step request, uses tools correctly instead of hallucinating their output, holds a longer document in context without losing the thread, and is far more reliable at the structured output that agents depend on.

Why it is behind the unlock

Because it is the honest thing to charge for. The free edition is not crippled: it is a complete assistant with chat, voice, tools, agents, documents and memory. What you buy is the ceiling.

The two ways to get it

In MyBenAI Lite it is part of the one-time $2.99 unlock, alongside vision, image generation and deep research. Every new Lite install also gets seven days of full access first — not a subscription trial, no card on file, nothing to cancel; when it ends the four features lock and the rest of the app carries on working free.

Or buy the paid app for $2, where all of it is included unconditionally and there is no paywall screen at all. That is the cheaper route, and it is meant to be.