Search results for "Large"

FlagEval (Libra) is a large-scale model evaluation system and open platform initiated by Beijing Zhiyuan Artificial Intelligence Research Institute, aiming to establish scientific, fair, and open evaluation benchmarks and methods. The platform has innovatively constructed a three-dimensional

CMMLU: Authoritative Chinese Language Model Knowledge Understanding Ability Evaluation Benchmark

109

CMMLU (Chinese Massive Multitask Language Understanding) is a large-scale language understanding benchmark designed specifically for the Chinese language context, covering 67 subject topics from beginner to advanced professional levels, including natural sciences, social sciences, engineering

2026-05-30

AGI Eval: A large model evaluation community and authoritative third-party evaluation platform

107

AGI Eval is a large model evaluation community jointly created by top universities and institutions such as Shanghai Jiao Tong University, Tongji University, East China Normal University, and DataWhale, with the mission of "assisting evaluation and making AI a better partner for humanity". The

OpenCompass: An Open Source Large Model Comprehensive Evaluation System and Sinan Open Platform

OpenCompass is an open-source large model evaluation system launched by Shanghai Artificial Intelligence Laboratory, providing one-stop evaluation services for large language models, multimodal models, and scientific intelligence models. The platform supports one click distributed evaluation of

MMLU: The International Authoritative Benchmark for Multi Task Language Understanding Ability of

MMLU (Massive Multitask Language Understanding) is a large-scale multi task language understanding evaluation dataset jointly released by the University of California, Berkeley and other institutions. It covers 57 disciplinary fields, including humanities, social sciences, natural sciences,

HELM: Stanford University led comprehensive evaluation framework for large language models and high

HELM (Holistic Evaluation of Language Models) is a comprehensive language model evaluation framework initiated by the Stanford University Center for Fundamental Model Research (CRFM), aimed at systematically evaluating large language models through multidimensional, standardized, and reproducible

2026-05-30

LangGPT - Structured Prompt Word Design Framework: Natural Language Programming for Large Language

LangGPT is a structured prompt word design framework inspired by programming languages, aimed at solving the problem of non AI experts having difficulty writing high-quality prompt words. Drawing on the design ideas of object-oriented programming languages, LangGPT proposed a dual layer structure

2026-05-29

A Quick Take: Just how much hassle can that AI meeting note-taker hidden inside Baidu Netdisk really save you?

"Jiandan Tingji" is an AI-powered speech-to-text tool launched by Baidu Netdisk. Integrated with the ERNIE Bot (Wenxin Yiyan) large language model, it supports the transcription of meetings, interviews, and lectures, as well as the intelligent generation of meeting minutes, achieving an accuracy rate of up to 97%. It is available across all platforms, with a continuous monthly subscription priced at 25 yuan.

2026-06-29

StableLM—an open-source AI model from the same family as Stable Diffusion, capable of running with only 3B parameters and commercially viable.

StableLM is an open-source large language model series launched by Stability AI. It is trained on The Pile extended dataset, which contains 1.5 trillion tokens. It offers multiple parameter versions, including 1.6B, 3B, 7B, and 12B, and supports text generation and code writing. It is licensed under the CC BY-SA 4.0 open-source license, allowing free commercial use. Stable LM 2 12B outperforms Llama 2 70B in some benchmark tests. It is suitable for developers, researchers, and small and medium-sized enterprises for private deployment.

Alibaba's M6—the world's first 10 trillion-parameter multimodal large model—consumes only 1% of the energy of GPT-3.

M6 is a general-purpose multimodal large model developed by Alibaba DAMO Academy. In October 2021, it became the world's first AI pre-trained model with 10 trillion parameters. Based on the self-developed Whale distributed framework, it was trained in just 10 days using only 512 V100 GPUs, with energy consumption only 1% of GPT-3. It supports tasks such as image and text generation, visual question answering, and poetry creation, and has been deployed in more than 40 scenarios, including Tmall virtual anchors and Taobao search. It is the core predecessor of the Tongyi Thousand Questions large model.

LLaMA—Meta's open-source AI model revolution, with performance comparable to GPT-4 and completely free for commercial use.

LLaMA is an open-source family of large language models launched by Meta, evolving from LLaMA 1 in February 2023 to multimodal LLaMA 4. The 70B parameter version boasts performance close to GPT-4, while the 405B version rivals GPT-4o. It is completely open-source and commercially viable. Enterprises can download and deploy it on their own servers without data sharing, meeting compliance requirements in industries such as finance and healthcare. With over 1.2 billion downloads globally, it is suitable for developers and enterprises requiring private deployment and deep customization.

ZCOOL – China's inspiration garden for designers, where 18 million creatives find direction.

ZCOOL is China's largest interactive community for designers. Founded in 2006, it brings together 18 million designers, illustrators, photographers and other creative professionals.

2026-07-04

360AI Search – an AI search engine that can "think slowly" for you, so you no longer need to flip through dozens of pages to find information.

360AI Search (Nano AI Search) is an AI-native search engine launched by 360, supporting multimodal input including text, voice, photos, and videos. It features a built-in "slow thinking mode," capable of calling up three or more large models at once and generating a 5,000-word in-depth report after reading 250,000 documents. It supports extracting mind maps from 100-page PDF/Word documents in one minute, extracting key points from one hour of video in one minute, and providing a one-stop AI tool for webpage AI summarization, PPT generation, and academic source tracing. In July 2024, user visits exceeded 90 million, transforming search from "providing links" to "providing direct answers."

MiTa AI Search – a free AI search engine that thinks and provides references, so you no longer need to flip through hundreds of pages of materials when doing research.

MetaAI Search is an intelligent search engine launched by Shanghai MetaNet Technology Co., Ltd. based on its self-developed MetaLLM large-scale model. It offers three search modes: concise, in-depth, and research-oriented. It supports searches across the entire web, document libraries, academic resources, podcasts, and more. Search results automatically generate outlines and mind maps, with each result accompanied by references. It incorporates the DeepSeek R1 deep thinking model, supporting a "think first, then search" research mode. It is compatible with web pages, apps, and mini-programs, with a free daily search limit of 100 times. Suitable for students, researchers, and professionals, it transforms information retrieval from "providing links" to "providing direct answers."

360 Smart Drawing – an AI drawing tool equipped with a massive model containing hundreds of billions of parameters, capable of generating thousands of styles with a single click.

360 Smart Drawing is an AI image creation platform launched by 360 based on its self-developed "360 Brain" large-scale model with hundreds of billions of parameters. Integrating the 360CV large-scale model and multimodal technology, it provides full-scene AI drawing and retouching functions, including text-to-image, image-to-image, doodle-to-image, partial redrawing, AI portrait, and LoRA model training. It has built-in models for thousands of styles, including traditional Chinese style, realistic, animation, and 3D, and supports bilingual (Chinese and English) prompts. As a member of the 360 AI ecosystem, it has been integrated into the entire 360 platform ecosystem.

2026-07-04

Rabbit R1 – An AI hardware device that challenges the "App terminator," a small orange box that went from being ridiculed to gaining popularity.

The Rabbit R1 is an AI-native pocket assistant launched by Rabbit Inc. It features a Large Motion Model (LAM) and allows users to directly control apps and complete tasks via voice commands. Although it was controversial due to its incomplete functionality after its release in early 2024, its reputation has reversed after more than 35 OTA updates over two years and a rewrite of Rabbit OS 2. The company also plans to launch a more geeky Cyberdeck product.

2026-07-09

PubMedQA—an "AI benchmark" specifically designed for biomedical question answering; only by comprehending research papers can an AI truly pass the "Medical Turing Test."

PubMedQA is the first question-answering dataset requiring reasoning over biomedical research texts; it was released in 2019 by institutions including the University of Pittsburgh. The task involves answering "Yes," "No," or "Maybe" questions based on PubMed abstracts. Comprising 1,000 expert-annotated samples and 211,000 artificially generated ones, the dataset aims to evaluate the ability of AI models to comprehend and reason about complex medical literature. It is widely used to benchmark the performance of large language models in the medical domain.

2026-07-10

SuperCLUE—A benchmark for Chinese large models: from foundational capabilities to AI agents, a single test reveals who is "swimming naked."

SuperCLUE is a comprehensive evaluation benchmark for general-purpose Chinese large models, released by the CLUE team. It provides authoritative, multi-dimensional capability assessments and rankings for Chinese large models by regularly publishing monthly and semi-annual reports based on three key benchmarks—open-domain multi-turn dialogue, closed-domain objective questions, and anonymous head-to-head battles—as well as dimensions such as mathematical reasoning, code generation, and AI agents.

2026-07-10

AI News (8)

OpenAI's GPT-5.6 series preview remains in internal testing, with proprietary inference chips set for mass production by year-end.

123

OpenAI is officially advancing the limited beta testing of its GPT-5.6 model series and has announced a partnership with Broadcom to co-develop "Jalapeño," a dedicated inference ASIC. The project targets tape-out within nine months and large-scale deployment by the end of 2026, aiming to slash inference costs by 50% and effectively completing the company's vertically integrated strategy combining models and chips.

Anthropic's valuation has reached $965 billion, surpassing OpenAI for the first time.

118

The AI industry is undergoing a seismic shift as Anthropic’s post-investment valuation climbs to $965 billion, surpassing OpenAI for the first time. Its Claude Opus 4.8 model has reached the pinnacle of reasoning capabilities, while Claude Code captures over half the code-focused AI market with $6.3 billion in annual revenue; this analysis delves into the company's three core strengths: capital, technology, and commercialization.

Meituan Open-Sources LongCat-2.0, a 1.6-Trillion-Parameter MoE Model; Full-Stack Execution on Domestic Computing Hardware Successfully Achieved

Meituan has officially open-sourced LongCat-2.0, a massive MoE model with 1.6 trillion parameters. This release achieves end-to-end training and inference deployment on a domestic computing cluster comprising 50,000 GPUs, thereby breaking reliance on overseas computing power. The announcement details LongCat-2.0's technical highlights, performance advantages, and industry value.

The wave of mandatory retirement is coming! Codex launches GPT-5.2/5.3-Code in June, developers

How do you view Minimax's new M3 multimodal model and updated Token Plan?

I originally wanted to say 'M3 is very strong' or 'just like that' directly, but later I found out that it's not something that can be explained in a single sentence. Let me admit one thing first: I have a bias against Minimax. It's not the kind of malice, it's the prejudice of being 'tricked'.

Sun Zhengyi was the richest man in Asia again, and fell again three days later: how to understand

To be honest, my first reaction when I saw this news was not "wow", but "is it again?" - because the last time he was so successful was in 2017. After that, everyone knows the story: WeWork went bankrupt, SoftBank Vision Fund suffered huge losses, his personal wealth once shrank to $21.1 billion,

The independent variable robot releases the world's first "event level prediction" embodied

Breaking the limitations of traditional frame by frame learning, the way robots understand tasks has entered a new stage

Out of control and overturned! Google AI unauthorized tampering with production code, resulting in