Research
Researchers release open-source sparse attention method for ultra-long context
The team says sparsified attention on Qwen 3.8 27B shows no observed accuracy loss, with the code, SGLang integration and paper all open. Accuracy drops on some tasks at 10M-token context, though needle-in-a-haystack mostly holds.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead