HiLS-Attention 7B Released

pmttyji · reddit · 2026-07-10

**Tencent/HiLS-Attention-7B** has been released on Hugging Face, alongside the paper "**Hierarchical Sparse Attention Done Right: Toward Infinite Context Modeling**." This is a continued training checkpoint based on **OLMo3-7B**, focusing on efficient long-context modeling. The core ideas proposed in the paper are: - Traditional block sparse attention requires computing full chunk mass upfront, which is computationally expensive - HiLS-Attention uses **compressed chunk keys** to estimate chunk quality - It then decomposes attention into a two-level softmax: **inter-chunk** and **intra-chunk** - This allows sparse attention to be learned end-to-end under **next-token prediction loss** The repo also notes that this is a **pretrained base model** without alignment or safety fine-tuning, meaning it may contain biases and produce unsafe content.

Related event: Tencent Releases HiLS-Attention 7B Model(2 posts)→

Original post →

More from Infra

Infra channel →