> VibeThinker-1.5B is a 1.5‑billion‑parameter dense language model from WeiboAI, fine‑tuned from Qwen2.5-Math-1.5B and purpose‑built for competitive math and algorithmic coding problems, where it delivers frontier‑level reasoning despite its small size. Trained with the “Spectrum‑to‑Signal Principle” framework that first maximizes solution diversity in supervised fine‑tuning and then reinforces correct reasoning paths via reinforcement learning, it achieves scores of 80.3, 74.4, and 50.4 on AIME24, AIME25, and HMMT25 respectively—surpassing the much larger DeepSeek R1—and reaches 55.9/51.1 on LiveCodeBench v5/v6, rivaling or beating larger models like Magistral Medium. The model is fully open source under the MIT license, trained with a reported post‑training cost of about $7,800, and is recommended for use specifically on math and coding tasks with long output lengths (up to around 40k tokens) using moderate sampling temperatures. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):
Модальности:
Генерация текста
Области применения:
Генерация кода Математика Диалог / чат
Задача: Генерация текста
Автор: prithivMLmods
Теги: gguf, qwen2, gpqa, text-generation-inference, math, code, en, endpoints_compatible
Лайков: 3 | Загрузок: 1,511
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
