> Nanbeige4.1-3B from Nanbeige is a compact 3B-parameter decoder-only Transformer language model (both Base and Thinking variants) pre-trained on 23T high-quality tokens using hybrid filtering and WSD strategies, followed by multi-stage post-training—30M+ SFT samples, thoughtfulT refinement, dual-level distillation from larger Nanbeige models, and RL—achieving state-of-the-art small-model reasoning that outperforms Qwen3-8B/30B-32B on AIME2024/2025 (SOTA averages), GPQA-Diamo Designed for deep single-pass multi-step reasoning on math/science/coding/puzzles without agentic loops, it employs Fine-Grained Warmup-Stable-Decay scheduling (0.1T warmup + 18.9T stable phases shifting to top-quality data) for superior token/sequence-level performance, matching 10x-larger models on demanding tasks while enabling consumer-grade local deployment under open license. (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow…
Модальности:
Генерация текста
Области применения:
Диалог / чат
Задача: Генерация текста
Автор: prithivMLmods
Теги: gguf, llama, text-generation-inference, en, endpoints_compatible, conversational
Лайков: 4 | Загрузок: 125
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
