Orvena 4B: the on-device model trained for what you ask your iPhone to do

Orvena 4B is the language model that Orvena runs on your iPhone from version 2.5. It is Orvena's own model, trained for the things people ask an assistant on their phone to do: move a meeting, set an alarm for when to leave, find a place and how long it takes to get there, search the web, or work in Notion, Todoist or Airtable. It runs on the phone, answers in 30 languages, and is free in Orvena with no account. This guide covers what it was trained for, how it did on 547 test tasks, how fast it is on an iPhone, how to switch to it, and where the open weights are.

Orvena 4B at a glance

What it was trained to do

Orvena 4B learned from requests played out on a simulated iPhone that has Orvena's real tools: calendar, reminders, alarms and timers, places and travel times, weather, web search and reading pages, photos, Health, contacts, files, Music and scheduled tasks. The simulated phone also had MCP servers with the tool lists of Notion, Linear, GitHub, Todoist, Airtable, Atlassian, Gmail, Google Drive, Google Calendar, Granola, Home Assistant and Zapier.

Each scenario came with the phone's data, such as the calendar, contacts and places, and with checks on which tools the assistant should call and what its answer must contain. Only answers that passed the checks were kept, 3,397 scenarios in all. Two smaller sets teach restraint: requests no tool can do, and everyday questions asked while connections are installed, so it answers or says what it can't do instead of reaching for a tool. A third set keeps it following formatting instructions, such as "answer in three bullet points".

How it did on 547 phone tasks

Orvena tests models on its own bench: 547 tasks in 30 languages, none of them used in training. Each one runs through Orvena's agent loop on simulated phone data and passes only when the right tools are called with the right details and the answer says what it should, in the user's language. The scenarios and every result are published as Orvena/orvena-bench. Qwen 3.5 4B, the model Orvena downloaded before, ran with thinking off, like Orvena 4B.

Multi-step tasks changed the most, 43 of 70 against 17. "I've got that job interview tomorrow, set me an alarm for when I need to leave, I'm going on the tube" needs the calendar, the current location, the transit time and then the alarm. "My notes from the flat viewing have the plumber's number in them somewhere, remind me tomorrow at 10 to call him" needs a file search before the reminder can be written.

Connections don't distract it either. The bench asks the same 100 everyday requests twice, once on a phone with no connections and once with 2 to 6 installed. Orvena 4B passes 81 and 83, and Qwen 3.5 4B 64 and 60. One task in five still fails somewhere, and it sometimes acts where asking you for a missing detail would be better.

How fast it is on an iPhone

On an iPhone 17 Pro Max with Orvena, 14 tasks ran twice each with the phone cooled between runs: tool chains, requests missing a detail, and requests it can't do.

Thinking automatic was Orvena's default before Orvena 4B, and most of the difference is the thinking it no longer needs: it writes fewer tokens for the same task. Decoding runs at about 21 tokens per second on this phone, so writing less is what makes it faster.

How to switch to Orvena 4B

On a new install, the model page of setup recommends it: "Orvena 4B, Orvena's own model, is the best fit", with the download size. If you already use Orvena:

1. Update Orvena from the App Store. A "Meet Orvena 4B" screen appears once.
2. Tap "Switch to Orvena 4B". The screen gives the download size and says your chats and settings stay as they are.
3. Keep using Orvena while it downloads. A small ring in the chat header shows the progress, and when the model is ready a banner says "Now using Orvena 4B. You can switch back in Models."

You can also switch at any time: open Settings from the top-left menu, then Models, tap Get on Orvena 4B, and tap Use once it has downloaded. If you answer with Private Cloud Compute or your own cloud model, the screen only offers "Got it", and Orvena 4B waits in Models for when you want an on-device model.

Orvena 4B starts with thinking off. To let it reason before answering, set Thinking to Automatic or Custom in Models. Orvena remembers the setting for each model, so going back to Qwen 3.5 4B restores its own. Qwen 3.5 4B stays in Models, and on iOS 27 so does Apple Intelligence.

The open weights

Orvena 4B is published under Apache 2.0. Orvena/Orvena-4B holds the full-precision weights, a multi-token prediction head for faster decoding on servers, and the technical report with the data, training and evaluation in full. Orvena/Orvena-4B-MLX-4bit is the exact build the app downloads. It keeps the architecture of the model it was built from, Qwen3.5-4B, so runtimes that support that model, such as vLLM, SGLang, Transformers and MLX, run it as it is.

Common questions

Is Orvena 4B free?

Yes. It is free in Orvena with no account, no limits and no subscription. Premium Unlock, a one-time purchase, adds voice conversations, connections and scheduled tasks, and they all work with Orvena 4B.

Does Orvena 4B work offline?

Yes. Conversation, photos and the phone's own tools work with no connection. Web search, weather, places, travel times and connected services still need the internet, as the offline guide lists.

Which iPhones can run Orvena 4B?

iPhone 15 Pro and newer. The model needs a little over 3 GB of memory while it runs, and only phones with 8 GB leave enough for iOS. Older iPhones have 6 GB or less.

Can I go back to Qwen 3.5 4B?

Yes. Qwen 3.5 4B stays in Models. Tap Use on it, and Orvena switches back with its own thinking setting.

Is Orvena 4B open source?

Its weights are open under Apache 2.0, so anyone can download, run, fine-tune and ship them. The 547-task bench it was tested on is published too, under CC BY 4.0. The training data is not public.