https://huggingface.co/Qwen/Qwen3-ASR-1.7B

[2601.20614] Harder Is Better: Boosting Mathematical Reasoning via Difficulty-Aware GRPO and Multi-Aspect Question Reformulation…

[2601.22149] DynaWeb: Model-Based Reinforcement Learning of Web Agents https://arxiv.org/abs/2601.22149

еще на 100к семплов потюнел)

недавно вышел https://huggingface.co/microsoft/VibeVoice-ASR но он такой огромный. Гигу можно на cpu крутить. vibe voice еще не тестил

ASR https://github.com/salute-developers/GigaAM