A Model Catalog You Can Override

MyBenAI treats models as data, not as part of the app. The list of what exists, what each one needs, and which tier it suits is a catalog the app refreshes over the air — so when a better small model is released, it can reach your phone without an app update, a store review, or you doing anything.

What is in it

Around fifty entries across five kinds: chat models from 135M up to 9B (Qwen3 and Qwen3.5, Llama 3.2, Gemma, Phi-4 Mini, SmolLM3, LFM2), vision-language models, Whisper models from tiny to large-v3-turbo, embedding models for document search, and Stable Diffusion checkpoints. Each entry carries its download URL, file size, SHA-256, estimated runtime RAM, default context length, minimum device tier, and whether it can call tools natively.

The manager

Settings → Models shows what you have installed, how much storage it is using, which model is active, and when each was last used — with a delete button next to every one. Anything else the catalog offers that your phone can run is one tap away. If you would rather run a 4B model on a phone the app rated for 3B, that is your call to make.

Why the catalog is hosted, not bundled

A bundled list goes stale the day it ships. This one is fetched from mybenai.robomiri.com, and the request is recorded in the network log like every other — it is one of the small number of things the app ever asks the internet for. A bundled copy also ships inside the app, so a phone that has never been online still knows what it can run.

Uncensored variants

The catalog includes abliterated builds of several models for people who want an assistant that does not decline. They are listed plainly rather than hidden, they are the same architecture as their siblings, and they are entirely your choice to install.