Tag: #embeddings36 results
TTE-Flash introduces 'latent think tokens' to replace explicit Chain-of-Thought generation, significantly reducing computational overhead for multimodal embeddings. The model achieves superior performance on the MMEB-v2 benchmark while maintaining constant inference costs.
IBM has released the Granite Embedding Multilingual R2, an open-source model under the Apache 2.0 license. It offers state-of-the-art retrieval quality for sub-100M parameter models and supports a 32K context window.
Redis research shows fine-tuning embedding models for compositional sensitivity cuts dense retrieval generalization by up to 40%. This threatens agentic AI pipelines, where retrieval errors cascade into wrong actions. Semantic similarity fails to capture exact intent due to vector space tradeoffs.
OpenClaw 2026.4.15 updates default Anthropic models to Claude Opus 4.7 with image understanding and adds Gemini TTS support in the Google plugin. New UI features include a Model Auth status card, while memory enhancements bring cloud storage to LanceDB and GitHub Copilot embeddings. Fixes improve tool security and agent reliability.
20M+ Indian court cases with structured metadata, citation graphs (followed/distinguished/overruled), Voyage AI embeddings + BM25. Covers Supreme/High Courts/Tribunals; useful for legal NLP, GNNs, RAG eval, Indian lang models. API/bulk JSON/Parquet export, public domain.
Hugging Face announces multimodal embedding and reranker models integrated with Sentence Transformers. These models enable processing of text and image data for enhanced retrieval and ranking tasks. They are now available on the Hugging Face Hub for immediate use.
OpenClaw v2026.3.24 introduces OpenAI-compatible /v1/models and /v1/embeddings endpoints through its Gateway. The Control UI gets a major tools visibility overhaul with real-time availability checks.
VLouvain reformulates Louvain community detection to operate directly on embedding matrices using community vector sums, avoiding O(n^2) graph edges. It scales to 1.57M nodes (d=200) in 11,300s, outperforming cuGraph and others. As GraphRAG drop-in, it cuts indexing from 3hrs to 5.3min and lifts recall from 37.9% to 48.8%.
Arc Institute's Evo2 genomic foundation model, trained on 9.3T nucleotides, captures regulatory similarities between genes like VIM and DES that BLAST misses. Embeddings show high cosine similarity (0.948) for muscle tissue promoters despite no sequence match. This indicates Evo2 learns deeper biological patterns amid noisy signals.
Tested 14 embedding models on 15 Thai MTEB tasks using A100 GPUs on Thailand's LANTA supercomputer. Qwen3-Embedding-4B leads with 74.41 score, followed closely by KaLM-Gemma3-12B at 73.92. Results merged into official MTEB repo with interactive leaderboard.