# Joint Subspace Sparse Attention speeds up long-context LLM decoding

> The new JSSA method can decode up to 6.4× faster on 128K token contexts, while keeping accuracy near full attention.

Oossa · 2026-10-15 · https://oossa.com/en/joint-subspace-sparse-attention-speeds-up-long-context-llm-decoding

Researchers led by Jiwon Song released a technique called Joint Subspace Sparse Attention (JSSA). It selects tokens at the individual level without gathering them first, keeping the KV-cache traffic low. On an NVIDIA B200 GPU, JSSA reached up to 6.43× the decoding speed of full‑KV attention for 128 K‑token contexts. Accuracy stayed close to full‑KV and to model‑native dense‑sparse attention, even for models as large as 700 B parameters.

## The facts

- Published: Thu Oct 15 2026
- Speedup: up to 6.43× on a 128K‑token context on an NVIDIA B200 GPU

## Why it matters

Faster long‑context decoding lets developers build cheaper, more responsive AI assistants that can read and respond to very large documents.

## Sources & references

1. [Joint Subspace Sparse Attention for Efficient Long-Context Decoding](https://arxiv.org/abs/2610.19026) – arXiv, 2026-10-15

Last updated: 2026-10-15