Models

The SORAEN model family

We build models for different kinds of work, starting with SORAEN Lite. Each one is released with open weights and a public evaluation.

Training · evaluation in progress Open weights at release Text, images and voice in

SORAEN Lite

Our first model: a capable 9B all-rounder for writing, reasoning, code, images and voice, with a particular strength in answering from your documents. Small enough to run on a single GPU.

Request access How it's trained

Parameters: 9 billion
Built on: Qwen3.5-9B (Apache 2.0)
Training: Supervised fine-tuning on ~12,500 curated examples, then preference training (DPO)
Inputs: Text, images (photos, screenshots, scans), voice via speech-to-text
Strengths: Grounded answers with citations, abstaining when unsure, tool calling, instructions, many languages
Safety: Safety training, red-team benchmark, moderation on every API call
Licence: Open weights, Apache 2.0
Release: When it passes its evaluation

Roadmap

What comes next

Planned

SORAEN Pro

A larger model trained with the same recipe and judged by the same public tests: stronger reasoning, coding and long documents, for teams that need the most capable answers.

Planned

Reasoning mode

Step-by-step thinking for hard problems, switchable per request, so simple questions stay fast and cheap.

Planned

Specialised models

Separate models for code, vision and speech, each measured on tests built for that kind of work.

Planned

Natural voice

Spoken conversation with SORAEN, beyond typing or dictating a question.

How we release

Every model, the same promises

Clean data

Only data whose licence allows commercial use, with every source credited.

Public evaluation

Scores against the base model on the same tests, published with each release, including losses.

Safety tested

A red-team benchmark on every release: it must be no less safe than its base, and still answer harmless questions.

Open weights

Download and run it yourself, or use it through the SORAEN API and chat.