Latest

The result makes late-interaction embeddings practical for large-scale retrieval, since storage costs were the main barrier to adoption. Teams running vector search can cut memory and infrastructure spend dramatically while keeping near-baseline accuracy, provided their pipeline can retain originals for rescoring. Engineers comparing against Matryoshka-style truncation or binary quantization now have a concrete 21× datapoint.

Qdrant@qdrant_engineSource

Per-task pricing under US$0.09 on ARC-AGI-1 and ARC-AGI-2 makes Haiku 5.5 one of the cheaper options to clear the 61% mark on ARC-AGI-2, a benchmark where many larger models struggle. Teams running agent or reasoning workloads at scale can use these numbers to weigh small-model cost against accuracy. Comparisons with other labs' verified scores are available directly from ARC Prize.

Official accountARC Prize@arcprizeSource

Trading deliberate reasoning for a fast single-pass design makes video QA over 100x quicker, which matters for latency-sensitive applications like live video analysis. Builders working on video agents or segmentation can grab the 4B or 35B checkpoints and benchmark against their existing backbones using the reported mIoU numbers.

Aran Komatsuzaki (论文推介)@arankomatsuzakiSource

The 2B active-parameter design targets efficiency over raw capability, so the interesting question is how it trades benchmark rank for cheaper inference against larger dense models. Developers running edge or high-throughput deployments can evaluate the weights directly from the release link. The comparison with Qwen gives a concrete starting point for choosing between the two families.

Gorden Sun@Gorden_SunSource

The compute-efficiency framing is the unusual part: TasteVal scores models by compute needed versus human experts, not task accuracy alone. That makes it a leading indicator for when automated research could become economical, useful for labs and investors tracking self-improvement timelines. The 1/30 cost figure gives a concrete benchmark for comparing frontier models on research ability rather than chat performance.

kimmonismus@kimmonismusSource

The core shift is system-level: LES moves the bottleneck from memory capacity to flash bandwidth and expert-prediction hit rate, letting a 35B-class model live on consumer phones without a cloud round-trip. Developers building on-device apps can fine-tune with SRFT's 4:1 task-to-replay data mix while keeping general abilities largely intact. Weights and code are on Hugging Face and GitHub, so teams can experiment today.

shao__meng@shao__mengSource