In a bid to tackle instability in training large language models, DeepSeek researchers have revisited a mathematical algorithm from 1967 - the Sinkhorn-Knopp algorithm - to enhance and stabilize hyper connections in neural networks. Manifold Constr…
LLMRouter is a revolutionary system emerging from the University of Illinois Urbana-Champaign, built to address the challenge of choosing the best large language model (LLM) for any given task. Operating as a dynamic routing library, it decides whi…
Liquid AI has released its latest innovative checkpoint, the LFM2-2.6B-Exp, under its LFM2 family of models. This particular checkpoint introduces pure reinforcement learning on top of its pre-trained architecture, aiming to excel in instruction fo…
Have you ever wondered if an artificial intelligence system could mimic the human brain's way of organizing thoughts, memories, and connections? What if this system could link concepts, draw insights, and even “sleep” to consolidate its knowled…
InstaDeep has rolled out its groundbreaking Nucleotide Transformer v3 (NTv3), a multi-species genomics foundation model tailored for 1 MB context lengths at single-nucleotide resolution. Designed to elevate genomic prediction and design, NTv3 bridg…
Google Health AI has taken a significant step in revolutionizing medical transcription with the release of its MedASR model. Designed specifically for medical dictation, this speech-to-text model operates under the Conformer architecture and bridge…
Google DeepMind has introduced Gemma Scope 2, an innovative interpretability suite designed for the Gemma 3 language models. This tool helps researchers uncover the layers of AI reasoning, from its smallest (270M) to largest (27B) models. Addressin…
Meta AI has unveiled an innovative technology called Perception Encoder Audiovisual (PEAV). This new AI encoder bridges the gap between audio, video, and text, creating a unified system for tasks like classification and retrieval. Using a staggerin…
Google has taken a step forward in making AI-interactive designs more efficient by introducing A2UI, an open-source Agent-to-User Interface protocol. This system enables AI agents to create and deliver interactive, secure interfaces using a JSON fo…
Large language models (LLMs) are widely used for generating text, yet they face performance challenges as sequences grow. This blog post dives into KV caching, a transformative optimization method that accelerates token generation by reusing key-va…
Mistral AI introduces its latest Optical Character Recognition model, Mistral OCR 3, designed to revolutionize document AI handling. This compact yet powerful model excels in extracting interleaved text and images from various document formats whil…
Google recently introduced T5Gemma 2, a cutting-edge encoder-decoder model that builds upon the success of Gemma 3. This model is designed to handle multimodal inputs, such as images and text, using innovative techniques like SigLIP for image encod…
Meta AI has unveiled SAM Audio, a cutting-edge unified model designed to address one of the biggest challenges in audio editing: separating specific sounds from complex recordings. With the ability to isolate sounds like speech, music, or backgroun…
Thinking Machines Lab has introduced a groundbreaking update with the general availability of Tinker, their advanced training API designed to make large language model fine-tuning simpler and more practical. This release is highlighted by the inclu…
OpenAI has introduced a groundbreaking toolkit, named ‘circuit-sparsity,’ designed to connect weight-sparse models with dense baselines. This new release includes models trained during optimization rather than post hoc pruning, fostering a better u…