How do you design a single model that can listen, see, read and respond in real time across text, image, video and audio without losing the efficiency? Meituan’s LongCat team has released LongCat ...

Optical character recognition has moved from plain text extraction to document intelligence. Modern systems must read scanned and digital PDFs in one pass, preserve layout, detect tables, extract key ...

How can a small model learn to solve tasks it currently fails at, without rote imitation or relying on a correct rollout? A team of researchers from Google Cloud AI Research and UCLA have released a ...

Every Ling 2.0 model uses the same sparse Mixture of Experts layer. Each layer has 256 routed experts and one shared expert. The router picks 8 routed experts for every token, the shared expert is ...

How do you convert real agent traces into reinforcement learning RL transitions to improve policy LLMs without changing your existing agent stack? Microsoft AI team releases Agent Lightning to help ...

What is new in Granite 4.0 Nano series? Granite 4.0 Nano consists of four model lines and their base counterparts. Granite 4.0 H 1B uses a hybrid SSM based architecture and is about 1.5B parameters.

Can an open source MoE truly power agentic coding workflows at a fraction of flagship model costs while sustaining long-horizon tool use across MCP, shell, browser, retrieval, and code? MiniMax team ...

Can we render long texts as images and use a VLM to achieve 3–4× token compression, preserving accuracy while scaling a 128K context toward 1M-token workloads? A team of researchers from Zhipu AI ...

Max Tokens is the maximum number of tokens the model can generate during a run. The model will try to stay within this limit across all turns. If it exceeds the specified number, the run will stop and ...

AI companies use model specifications to define target behaviors during training and evaluation. Do current specs state the intended behaviors with enough precision, and do frontier models exhibit ...

In this article we will analyze how Google, OpenAI, and Anthropic are productizing ‘agentic’ capabilities across computer-use control, tool/function calling, orchestration, governance, and enterprise ...

Web agents often fail when layouts shift or when tasks require long sequences. WALT targets this failure mode by mining site functionality offline, then exposing it as tools that encapsulate ...