AI Briefing
KOSign in

Maincode releases Matilda Platypus, a family of Qwen 3 fine-tunes using Model Stitching

·2026.11.10 11:09

Key point

The 13B member of the family uses 91% of the larger parent's compute while maintaining a base model status.

Details

  • Australian AI lab Maincode released the Matilda Platypus family, a set of Qwen 3 fine-tunes created using Model Stitching.
  • The method connects lower layers of a larger pretrained Qwen3 model to upper layers of a smaller one, using a learned linear Representation Bridge to align hidden representations.
  • The 13B checkpoint (cut 29+13) contains 13.51B parameters and requires 26.3 GFLOP per token, achieving 91% of the larger parent's compute.
  • The models are base models without instruction tuning, ranging in size from 1.5B to 13B parameters.
  • A technical report is available for the method and full evaluation, with models hosted on Hugging Face.
fine-tuninghugging-faceqwen-3maincodemodel-stitchingmatilda-platypus
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

HuggingFace BlogHugging Face Releases Fine-Tuning Guide for TRELLIS.2 3D Generation ModelsRedditHugging Face Introduces Harness Adapters for Task-Specific LoRA TrainingRedditNex-AGI Releases Nex-N2.6: 321B Multimodal Agentic Model for Computer UsePyTorchKRVIDRAFT Publishes Leaderboard Grading Operational Forecasts DailyTLDR AIHugging Face Hack Reveals Frontier AI Alignment Failure, Not Cybersecurity GapTLDR AIRunway Releases Praxis-1 Open-Weight Robotics Model