Search

Date Range
Advanced
Adjust the alpha value (0 to 1)
Adjust the relevance threshold (0 to 1)
arxiv.org Scientific research

[2611.06131] ModalSearcher: Weaving Visual Evidence into Autonomous Cross-Modal Deep Search

The paper introduces autonomous cross-modal search for multimodal LLMs, enabling agents to switch among reasoning, text, and visual search across multi-hop tasks, and to interleave modalities within a knowledge graph. It reports a training pipeline and benchmarks showing ModalSearcher-8B outperforms a baseline on several multimodal QA tasks.

1 min read Listed 8 days ago in Bluesky
arxiv.org Documentation

[2610.39301] Youtu-Parsing-Omni: One Encoder, One Schema, All Modalities

A single 5B-parameter omni-modal model uses one encoder per modality and a unified schema to parse multiple modalities, with a self-evolving training loop (OSAD) that aligns outputs across tasks and modalities; it achieves state-of-the-art results on OmniDocBench and OmniParsingBench, and code and weights are public.

2 mins read Listed 17 days ago in Hugging Face Daily Papers
www.scriptbyai.com Other

AI Model Release Calendar: Latest AI Model Releases by Date

A continuously updated calendar lists major AI model releases by date, provider, and type. It explains how release dates are assigned and includes model launches across text, coding, multimodal, image, audio, and other fields.

9 mins read Listed 2 days ago in Brave News
arxiv.org Scientific research

[2610.32411] A Broader Look into MLLM's Retrieval Heads

The paper investigates retrieval heads in multimodal LLMs, showing they remain sparse and largely shared across modalities even with audio-visual inputs; ablating them harms multimodal tasks, while optimizing their attention improves retrieval metrics.

2 mins read Listed 23 days ago in Hugging Face Daily Papers
huggingface.co Tutorial

Paper page - WeVisOne-VL: Unified Visual Perception from Objects to Pixels

The document describes a user interface for uploading media (images, audio, video) via drag-and-drop, paste, or click, with a focus on a unified visual perception system called WeVisOne-VL that handles multiple input modalities.

1 min read Listed 22 days ago in Bluesky
www.unite.ai News

Liquid AI Launches LFM3 Family of Six Open-Weight 2B-Class Models – Unite.AI

Liquid AI unveils LFM3, a family of open-weight, multi-modal Liquid Foundation Models on a shared backbone, with six 2B-parameter variants handling text, vision, audio, omni input, and decision tasks. The release emphasizes a tri-hybrid architecture, broad hardware support, strong pre/post-training regimes, partner testing, and superior efficiency and performance versus prior models.

5 mins read Listed 18 days ago in Brave News
silentstars.dev Tools and repositories

Novera · SilentStars

Novera · SilentStars is a local/cloud AI workspace for multi-media tasks, with a 1.0.0 release across Windows, Linux, macOS, and Android, and broad language support; Mac builds lack Apple notarization. TRON-ABS is referenced as 'the art of Human with AI.'

1 min read Listed 10 days ago in Bluesky
huggingface.co Documentation

encoding/README.md · deepseek-ai/DeepSeek-V4.1-Flash-1017 at main

DeepSeek-V4.1 introduces a text-and-vision encoding reference, detailing prompt-format changes, reasoning budget, and system/multi-modal support. It outlines DSML syntax differences, thinking modes, and practical usage notes for encoding and parsing model interactions.

7 mins read Listed 28 days ago in reddit/huggingface/🐳 DeepSeek-V4.1-Flash-1017 released: MIT weights, 1M context, and stronger coding and agent benchmark results
arxiv.org Tools and repositories

[2610.27103] Omni-Woodpecker: Hallucination Correction via Evidence-Prior Rebalancing for Omni LLMs

A training-free framework, Omni-Woodpecker, reduces cross-modal hallucinations in Omni LLMs by rebalancing fused representations to align evidence with priors, showing strong cross-architecture performance and publicly available code.

1 min read Listed 23 days ago in Bluesky
www.howtogeek.com News

Mistral is taking on U.S. and Chinese AI giants with its biggest model yet

Mistral unveils Large 4, aka Le Chonk, a 1.05T-parameter open-weight model designed for coding, cybersecurity, and multi-modal tasks, with European deployment plans and a public preview that positions it as an alternative to US/Chinese closed models; it uses mixture-of-experts, supports a 1M-token context and 160+ languages, and faces policy-driven task limitations alongside notable benchmark scores.

4 mins read Listed 16 days ago in Google News