Generates L2-normalized multimodal embeddings (default 4,096‑D) for text, images, videos and visual documents, supporting interleaved inputs and flexible dimension truncation (Matryoshka). Designed for cross-modal retrieval, ranking and downstream retrieval systems; audio is not supported.
Generates unified embeddings for text, images, video, visual documents and interleaved multimodal inputs with configurable output dimensions and Matryoshka truncation to trade accuracy for cost. Model weights and code are released under Apache-2.0; the 9B variant scores 80.6 on MMEB-v2.
A 10‑billion‑document retrieval benchmark with per‑document 768‑dim unit‑norm dense embeddings and mGTE sparse embeddings, FineWeb text/metadata, and exact top‑1000 MS MARCO ground truth for ~120k queries. Built for large‑scale evaluation of dense/sparse/hybrid retrieval, filtered search, indexing, ANNS algorithms, and embedding compression.
Generates unified 768‑dimensional embeddings for text (including code), images, video and audio to enable cross‑modal semantic search and retrieval. Supports task instruction prefixes, Matryoshka truncation to 128/256/512/768 dims, and modular encoders for on‑device use under an Apache‑2.0 license.
Runs locally on constrained devices to turn text into guaranteed-parsable JSON tool calls, typed structured extractions, or sentence embeddings. Delivered as a single compact weights file (8–29 MB) with a laddered 2–20-layer design, low-bit quantisation and calibrated confidence scores for on-device apps.
Scores candidate actions against a textual state using contrastive state/action embeddings for very fast zero-shot ranking and typed decision-making. Built as two small projection heads on frozen Qwen3-8B; fine-tunable as a verifier for agentic benchmarks.
Synthetic, clinician-verified ChatML dataset of 2,194 doctor–patient encounters covering 2,194 unique human diseases; each JSONL record includes 20 structured fields, verified PubMed references, realistic vitals/labs, and is intended for RAG and model fine-tuning (not medical advice).
On-device speech-to-text for short clips (up to 30s) in seven languages, yielding transcripts, word-level timestamps and per-frame speech embeddings in a single 16.9 MB model. Runs on the Needle CPU engine with 2–4 bit quantization, supports keyword biasing and returns empty transcripts for silence.
GGUF-packaged weights for EmbeddingGemma 2 enabling local multimodal embeddings (text, image, video, audio); supports 768/512/256/128 dimensions, BF16/FP32 inference, selective modality loading for lower memory, and is suited for semantic search and RAG.
Injects corpus knowledge into a language model’s intermediate layers, avoiding explicit retrieval at inference. It predicts shared memory per text chunk and reduces the retrieval work needed to construct training targets while keeping the backbone frozen.
Proves that JEPAs across modalities can learn representations that differ only by a rotation when positive pairs describe the same world states. These “good views” achieve cross-modal alignment with far fewer samples than next-token prediction.