{"name":"latent-sparse-attention","repo":"Prefrontal-AI/latent-sparse-attention","url":"https://midorreal.com/project/prefrontal-ai-latent-sparse-attention","repository":"https://github.com/Prefrontal-AI/latent-sparse-attention","description":"Code for Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention","facts":{"license":{"value":"Apache-2.0","source":"https://github.com/Prefrontal-AI/latent-sparse-attention","checked":"21 Nov 2026"},"language":{"value":"Python","source":"https://github.com/Prefrontal-AI/latent-sparse-attention","checked":"21 Nov 2026"},"last_release":{"value":"qsa-weights-v1, 9 Nov 2026","source":"https://github.com/Prefrontal-AI/latent-sparse-attention/releases","checked":"21 Nov 2026"},"releases":{"value":"6","source":"https://github.com/Prefrontal-AI/latent-sparse-attention/releases","checked":"20 Nov 2026"},"contributors":{"value":"2","source":"https://github.com/Prefrontal-AI/latent-sparse-attention/graphs/contributors","checked":"20 Nov 2026"},"archived":{"value":"no","source":"https://github.com/Prefrontal-AI/latent-sparse-attention","checked":"21 Nov 2026"}},"stars":474.0,"writeup":{"what_it_is":"Code for Latent Sparse Attention (LSA), a method that clusters keys hierarchically in a learned latent space so each query scores clusters instead of every key. The README describes a distilled variant (D-LSA), a native variant trained from scratch (N-LSA), and an SGLang plugin for serving.","audience":"Researchers and engineers working on long-context attention in large language models.","claims":[{"kind":"specific","claim":"On GLM-5.3-Flash at 10M tokens, one attention layer is 22.7x faster in prefill and 4.6x faster in decode than the model's own sparse attention.","excerpt":"On GLM-5.3-Flash at 10M tokens it speeds up one attention layer by 22.7x in prefill and 4.6x in decode over the model's own sparse attention.","status":"not_checked"},{"kind":"specific","claim":"On Qwen3.8-27B, D-LSA is 72.4x and 42.5x faster than full attention.","excerpt":"On Qwen3.8-27B it is 72.4x and 42.5x faster than full attention.","status":"not_checked"},{"kind":"specific","claim":"N-LSA matches the accuracy of full attention, and its training is 1.6x faster per layer at 128k tokens.","excerpt":"It matches the accuracy of full attention, and its training is 1.6x faster per layer at 128k tokens.","status":"not_checked"},{"kind":"specific","claim":"The attention kernel is unchanged; LSA only decides which keys are read.","excerpt":"The attention kernel is unchanged.","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-20"},"why_now":null,"owner_supplied":null,"score":{"verdict":"sleeper","hype":30,"reality":67,"gap":-37,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-11-09T16:05:58+00:00","text":"Latest release qsa-weights-v1","source":"https://github.com/Prefrontal-AI/latent-sparse-attention/releases"},{"date":"2026-10-30T22:20:28+00:00","text":"First release","source":"https://github.com/Prefrontal-AI/latent-sparse-attention/releases"},{"date":"2026-10-27T09:52:28+00:00","text":"Repository created","source":"https://github.com/Prefrontal-AI/latent-sparse-attention"}],"cite":"latent-sparse-attention is a GitHub project at https://github.com/Prefrontal-AI/latent-sparse-attention written mainly in Python, described there as \"Code for Latent Sparse Attention: Towards Subquadratic Ultra-Long Context Attention\".","attribution":"Data from Mid or Real, https://midorreal.com/project/prefrontal-ai-latent-sparse-attention, checked 22 November 2026. Attribution with a link is required.","still_real":null}