Harrier's Decoder-Only Embeddings Hit SOTA Multilingual
Microsoft's open-source Harrier models (270M-27B params) top MTEB v2 benchmarks using decoder-only architecture, 32k context, and instruction prefixes—shifting embeddings toward LLM foundations while rivals cut video costs and add skills.
Decoder-Only Shift Powers Multilingual Retrieval
Microsoft's Harrier OSS v1 family—models at 270M, 600M, and 27B parameters—achieves state-of-the-art on MTEB v2 benchmark for classification, clustering, retrieval, and paraphrase across languages. Unlike BERT-style encoders, these use decoder-only architecture like modern LLMs: final representation from the last token, normalized for consistency. This enables 32,768-token context (vs. old 512-1k limits), processing full documents without chunking losses. For peak results, prefix queries with instructions (e.g., "retrieve semantically similar text") while encoding documents plain—boosts task-specific matching. Smaller models leverage knowledge distillation from larger ones for efficient deployment balancing speed, memory, and cost. Builders gain production-ready multilingual semantic search without proprietary lock-in.
Video Generation Costs Halved for High-Volume Apps
Google's Veo 3.1 Light matches Veo 3.1 Fast speed at <50% cost via Gemini API, supporting text-to-video/image-to-video in 16:9 landscape or 9:16 portrait up to 1080p resolution and 4/6/8-second durations (pricing scales by length). Veo 3.1 Fast pricing drops April 7th, enabling iterative apps where users generate/refine multiples without budget strain—key for mainstream adoption. Gemini tests 3D avatars from likeness uploads for image/video gen, Remy exam-prep learning mode, and skill support for modular instructions, signaling education/multimodal pushes pre-I/O.
Hardware Adoption, UI Experiments, and Skill Modularity
Meta's prescription Ray-Ban glasses (Blazer/Scriber Optics Gen 2, $499+) fit most Rx types with adjustable pads/tips, 8hr battery (48hr cased), 3k-pixel video—prioritizing daily wear over gadget novelty. New: hands-free meal logging, E2E-encrypted WhatsApp summaries (on-device), expanded US navigation; cements 76.1% 2025 smart glasses share. Anthropic tests Epitaxy UI in Claude Code (hotkeys, model/skill selection, animations) amid NPM misconfig leak exposing models like Capybara/Strudel. xAI builds Grok custom skills: name/desc/instruction sets, importable via zip/skill/md, as reusable blocks beyond March's 4-agent limit—mirroring Anthropic/OpenAI/Google trends for prompt modularity over one-offs.