Chubby♨️@kimmonismus · XZyphra发布ZAYA1-8B模型，其活跃参数不足10亿，却在数学、编程和推理基准测试中媲美更大的开源及专有系统。其亮点不仅在于小尺寸，更在于全栈技术方案：完全基于AMD基础设施训练，采用了新的架构选择和大规模强化学习。此外，模型应用了一种名为Markovian RSA的测试时计算方法，通过并行推理和递归聚合，显著提升了复杂数学问题的解决能力。
Zyphra under 1B active parameters， AMD-Trained， big evals， look strong？
Zyphra says its new ZAYA1-8B model delivers unusually high reasoning power for its size， using under 1 billion （！） active parameters while competing with much larger open-weight and proprietary systems on math， coding， and reasoning benchmarks.
The interesting part is not just the model's size， but its full-stack bet： AMD-only training infrastructure （！）， new architectural choices， large-scale RL， and a test-time compute method called Markovian RSA that appears to boost hard math performance through parallel reasoning and recursive aggregation.