Research | Data Craftsman

Research

Papers, essays and model releases from the lab. Everything we release is open: weights on Hugging Face, code on GitHub, and a model card that says plainly what the model can and can't do.

01 / Publications

Papers and essays

Technical reports for each model release will be published here.

MediumAI Agents: If They Were Perfect, What Would I Ask of Them?

Follow on Medium

02 / Released

Models you can download today

piper-en_AU-librivox-medium

Ten Australian English voices for Piper, the open source text to speech engine. Trained on public domain LibriVox recordings, small enough to run on a CPU, and fully offline.

~70M parameters
77 MB
Runs on CPU
CC BY 4.0

View on Hugging Face

Text to speech

03 / In development

Our model roadmap

We are training the recurrent language model now. Next comes an open Australian decision model, then a new generation of Australian voice technology. Our long term goal is to pretrain a language model from scratch under a new architecture once we secure the funding and GPU compute.

01 / 04

01 / Training now

Australian recurrent language model

A 26B mixture of experts model for long documents, code, tools and Australian workplace tasks.

Gemma 4
Recurrent depth
Australian English
Text + image
Open weights
Apache 2.0

Base model: google/gemma-4-26B-A4B-it
Architecture: Gemma mixture of experts. Recurrent depth must beat the native model at matched compute before it ships.
Precision: NVFP4 deployment build. We will publish the final mixed precision map after measurement
Built to handle: Long documents, code, tools and real Australian workplace tasks
Deployment: One 32 GB+ Blackwell class workstation GPU
Quality target: An independent evaluator must measure ≥20 on the AA Intelligence Index. Australian benchmarks cover lexis, spelling, culture and knowledge, and the model must retain general and multimodal capability.
At release: Open weights, per task results including failures, and a technical report covering data, training and trade offs

02 / Up next

Australian decision model

An open source decision model inspired by Jev, designed to reason with Australian context, institutions and logic.

Open source
Australian logic
Decision reasoning

Purpose: Make structured decisions using Australian assumptions and context rather than relying on overseas defaults
Approach: Build an openly available Australian counterpart to models such as Jev

03 / Planned

Next generation Australian voice model

A higher quality Australian voice model to succeed our first Piper release and move beyond its 24 hour training run.

Text to speech
Australian English
Open source
Piper successor

Starting point: Lessons from piper-en_AU-librivox-medium, our first voice model, trained for 24 hours
Goal: A next generation voice system with more voices, a larger training effort, better voice quality and potentially different base

04 / Seeking compute

From scratch language model

A new language model pretrained from the ground up under a new architecture. This is the long term project, and it depends on funding and access to substantial GPU compute.

Pretraining
New architecture
From scratch
Funding required
GPU compute required

Research direction: Design and pretrain a language model rather than adapting an existing base model
Dependency: External funding, GPU hardware or cloud compute sufficient for pretraining
Status: Work on the architecture and training programme will begin when we secure the required resources

Contribute compute, evaluation data or early testing

04 / Open source

Follow the work as it happens

We publish new releases on Hugging Face first. Training and evaluation code lives on GitHub.

Hugging Face GitHub

Help train the next model