Polymathic AI, Princeton, and USC ISI researchers released FlashMap, a flow map drafter for speculative decoding that jointly drafts 16 tokens in a single pass, achieving up to 6.14x lossless speedup on Qwen3-8B (MarkTechPost). The method targets efficient inference on smaller models, a priority for on-device and resource-constrained deployments.
Liquid AI released LFM3, a suite of six open-weight 2B-class multimodal models covering text, vision, audio, omni, and decision-making modalities (MarkTechPost). The LFM3-1.7B variant features a tri-hybrid backbone, 256K context window, and quantizes to approximately 1 GB at 4-bit precision (Hugging Face Blog). In parallel, UC Berkeley’s LEANN vector index achieved 90% recall accuracy while compressing a 60-million-chunk RAG index from 201 GB to 6 GB by storing a pruned graph and recomputing embeddings at query time, though with seconds-level search latency as a trade-off (MarkTechPost).