Multilingual 3.7B MoE trained from scratch on a consumer GPU
Significant_Focus134 · reddit · 2026-09-01
The author released a checkpoint of 'Polanka 3.7B', a multilingual MoE model trained from scratch entirely on a single RTX 4090 over several months. It supports 13 languages (including Chinese and English) with upscaled data for PL/EN/ZH, and is shared as a research artifact.
More from Research
- Stop building remote-controlled robots; sim2real is the key to success — _Stocko_ · 2026-09-02
- CS Professor shares the evolution of their paper-reading stack over the years — CSProfKGD · 2026-09-02
- OpenBind preprint released with new virtual screening benchmark — MoAlQuraishi · 2026-09-02
- Anima Anandkumar launches acceleratedU: neural operators over transformers for physical prediction — AnimaAnandkumar · 2026-09-02
- New blog 'Proofs and Prompts' explores how AI is reshaping mathematics — giannis_daras · 2026-09-02
- EMNLP 2026 Paper: SCALE uses structured case law for legal reasoning — ponguru · 2026-09-02