#1Eagle 3.1: Collaboration Between the EAGLE Team, vLLM Team, and TorchSpec Team
Eagle 3.1 is a major speculative decoding breakthrough for LLM inference, tackling "attention drift" that destabilizes drafting at deeper speculation depths. The collaboration between the EAGLE research team, vLLM production inference team, and TorchSpec training infrastructure team delivers 2x longer acceptance lengths in long-context workloads and over 2x per-user output throughput at low concurrency. It's backward-compatible with Eagle 3 checkpoints, making deployment seamless across existing infrastructure.


