Daily AI News

Meituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention

Meituan unveiled LongCat-2.0, a 1.6‑trillion‑parameter mixture‑of‑experts model that activates roughly 48 billion parameters per token. The architecture introduces LongCat Sparse Attention and supports a native one‑million token context window, running on domestic AI ASIC superpods. It is positioned as a high‑throughput, open‑weight alternative for large‑scale language tasks.

Read source →

Sakana AI Launches Sakana Translate, a Namazu‑Powered Japanese–English–Chinese Translation Tool

Sakana AI added Sakana Translate to its Sakana Chat platform, leveraging the Namazu model series for bidirectional translation among Japanese, English, and Chinese. The tool offers three modes—Translate, Proofread, and Ask—and is evaluated on the XCOMET‑XL benchmark. It aims to streamline multilingual workflows for both consumers and enterprises.

Read source →

Synthetic Sciences Releases OpenScience: An Open‑Source, Model‑Agnostic AI Workbench for Scientific Research

OpenScience is a new Apache‑2.0 licensed workbench that integrates any frontier or open‑weight model via user‑provided API keys. It supports end‑to‑end pipelines across machine learning, biology, physics, and chemistry, and ships with over 250 editable skills. The platform is designed to accelerate interdisciplinary AI‑driven research.

Read source →

Baidu's "Unlimited OCR" Processes Dozens of Document Pages in One Pass by Treating Memory Like Human Forgetting

Baidu unveiled Unlimited OCR, a new optical‑character‑recognition system that can ingest dozens of pages in a single inference pass. The model uses a modified attention mechanism that keeps memory usage constant, mimicking human forgetting. It now leads the most important OCR benchmark by a significant margin.

Read source →

🤗 Kernels: Major Updates

Hugging Face announced a substantial overhaul of its 🤗 Kernels platform, adding support for GPU‑accelerated notebooks, collaborative editing, and seamless integration with the latest transformer libraries. The update also introduces built‑in benchmarking tools and a marketplace for reusable code snippets. These changes aim to accelerate rapid prototyping for AI developers.

Read source →

LlamaIndex ‘legal‑kb’: Agentic Retrieval over Index v2 with Retrieve, Find, Read, and Grep Tools

LlamaIndex released a public reference app called legal‑kb that provides agents with filesystem‑style access to a document knowledge base on Index v2. The app bundles retrieve, find, read, and grep capabilities with automatic versioning and visual citations, built on TanStack Start and AI SDKs. It targets legal professionals needing precise, auditable retrieval.

Read source →

AI Search Agents Don’t Fail at Searching, They Fail at Asking the Right Questions When Queries Get Ambiguous

A new benchmark called DiscoBench evaluates how AI agents handle ambiguous queries, revealing that many agents continue searching without prompting for clarification. The study highlights a critical gap in interactive search: the ability to ask follow‑up questions. Recommendations include integrating uncertainty estimation and dialogue‑driven refinement.

Read source →

Training Gemma‑3 for Structured Mathematical Reasoning with Tunix GRPO, LoRA Adapters, and GSM8K Rewards

Researchers present a pipeline that fine‑tunes Gemma‑3 using GRPO reinforcement learning, LoRA adapters, and a reward model based on GSM8K performance. The workflow improves the model’s ability to generate step‑by‑step mathematical solutions while maintaining format compliance. Results show a measurable boost on standard math benchmarks.

Read source →

AI Poses ‘Hiroshima’-Style Threat to Humanity Without Global Rules, Says Cooper

British Foreign Secretary Yvette Cooper warned that AI development, if left unregulated, could pose an existential risk comparable to nuclear weapons. She called for an international treaty to govern advanced AI research and deployment. The statement follows mounting concerns about autonomous weaponization and societal disruption.

Read source →

Amazon Will Stop Accepting New Customers for Mechanical Turk

Amazon announced that its crowdsourcing platform Mechanical Turk will no longer onboarding new requesters, signaling the service’s gradual phase‑out. Existing customers can continue using the platform, but no new projects will be launched after the cutoff. The move reflects shifting demand toward automated AI labeling pipelines.

Read source →