Thinking Machines 发布 Inkling 2 Small：594B 模型仅 68 小时完成预训练
Thinking Machines 把 594B 模型的预训练压到 68 小时，靠的是从第一步就做原生 4-bit 训练。这篇博客把量化误差怎么干扰学习、怎么解决讲得很透，做训练优化的工程师应该看看。
Thinking Machines 发布 Inkling 2 Small，这是一个 594B 参数的模型，在 GB300 NVL72 机架上用 NVIDIA NVFP4 低精度格式训练，预训练只花了 68 小时。NVIDIA 官方账号转发了这一成果。团队在博客 Connectionism 中披露了完整训练细节：从预训练第一步到最后的 RL rollout，全程使用原生 4-bit 前向传播，而不是先高精度训练再量化。
技术上的关键难点在于：比特数更少意味着训练更快，但量化误差会干扰学习过程。团队的方法是让模型从第一步起就适应 4-bit 精度，并解决了量化误差与梯度学习相互干扰的问题，使低精度训练能扩展到 594B 这样的规模。
对行业而言，NVFP4 加 GB300 NVL72 的组合展示了 4-bit 原生训练在超大规模模型上的可行性，若训练成本能按此量级压缩，中小团队训练大模型的门槛会明显降低。
Congrats @thinkymachines team on Inkling 2 Small! A 594B model pretrained in just 68 hours with NVIDIA NVFP4 on GB300 NVL72 racks. Read how the team made low-precision training work at this scale. 👇 Thinking Machines @thinkymachines Training with fewer bits is faster, but quantization error can interfere with learning. Today on Connectionism, we detail how we trained Inkling 2 Small with a native 4-bit forward pass from the first step of pretraining through the final RL rollout. thinkingmachines.ai/blog/end-to-en… 🔗 View Quoted Tweet📊 9 ⚡
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include Inkling 2 Small, Thinking Machines, NVIDIA, NVFP4.
Source attribution: curated from NVIDIA AI (@NVIDIAAI).
