EAGLE 3.1 tackles attention drift in speculative decoding with two architectural fixes, delivering up to 2x longer acceptance length on long-context workloads. The update ships with training support via TorchSpec and native vLLM integration.
May 27, 2026
Two fixes worth noting today. EAGLE 3.1 addresses attention drift in speculative decoding, a subtle production issue that was limiting acceptance length on long-context workloads. The patch delivers up to 2x improvement and ships with vLLM integration out of the box. Separately, Mem0's Node SDK v3.0.4 closes a silent bug where PGVector filter operators were returning wrong results across search and list operations, the kind of correctness issue that is easy to miss until it quietly corrupts your agent's memory retrieval.
EAGLE 3.1 tackles attention drift in speculative decoding with two architectural fixes, delivering up to 2x longer acceptance length on long-context workloads. The update ships with training support via TorchSpec and native vLLM integration.
Firecrawl is now available on the Vercel Marketplace, letting teams feed LLM-ready web data into AI agents without running their own crawling infrastructure. Key capabilities cover scraping, full-page search and retrieval, and interaction with dynamic pages.
CopilotKit v1.58.0 ships pre-built chat components for React Native, ending its status as a headless-only target. Mobile builders can now drop in the same chat surface the web SDK provides.
Mem0's Node SDK v3.0.4 patches a bug where PGVector filter operators were passed as raw values, returning incorrect results. All major filter operators now work correctly in search, keyword search, and list operations.