Maincode released a family of Qwen 3 base models made by joining layers from larger and smaller models. The family ranges from 1.5B to 13B parameters, according to a community post citing the model card.
Key points
The models are base models without instruction tuning
The family ranges from 1.5B to 13B parameters
The 13B checkpoint has 13.51B parameters and uses 91% of its larger parent’s compute
The LocalLLaMA community post reports that Australian AI lab Maincode released the Matilda Platypus models. The post says they were built by connecting lower layers from a larger pretrained Qwen 3 model to upper layers from a smaller one, with a learned linear bridge aligning their representations.
The post identifies the 13B model as the 29+13 layer combination and says it requires 26.3 GFLOP per token. It points to a technical report for the method and full evaluation; the post does not provide evaluation results.