Ulam发布3B数学推理模型Ulam-1-Small，对标百倍参数模型
Ulam-1-Small is here.
Ulam-1-Small is here.
Ulam-1-Small 来了。
Our first model is a 3B-parameter research-math LLM that matches some models 100x its size on research-level problems (like ErdosBench).
我们的首个模型是一个 3B 参数的研究级数学 LLM，在研究级问题（如 ErdosBench）上，其表现可与某些比它大 100 倍的模型相媲美。
It’s a post-trained (RL + SFT) version of VibeThinker-3B - already great at olympiad-level problems. We asked a simple question: can small models actually be useful for real research? The answer is yes.
这是 VibeThinker-3B 的后训练（RL + SFT）版本——它已经擅长奥林匹克级问题。我们问了一个简单的问题：小模型真的能对实际研究有用吗？答案是肯定的。
This was also a stress test of the datasets and problem sets we’re building at Ulam. Turns out even a “seemingly optimized” model still had room to level up.
这也是对我们正在 Ulam 构建的数据集和问题集的一次压力测试。事实证明，即使是“看似已优化”的模型仍有提升空间。
This is only the beginning. More releases are coming. Any feedback is welcome!
这仅仅是开始。更多版本即将发布。欢迎任何反馈！
HuggingFace weights: https://huggingface.co/ulamai/Ulam-1-Small
HuggingFace 权重：https://huggingface.co/ulamai/Ulam-1-Small
Learn more about our training: https://huggingface.co/ulamai/Ulam-1-Small/blob/main/paper/Ulam-1-Small-whitepaper.pdf
了解更多关于我们的训练：https://huggingface.co/ulamai/Ulam-1-Small/blob/main/paper/Ulam-1-Small-whitepaper.pdf
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力