Kwai has open‑sourced its new flagship multimodal model Keye‑VL‑671B‑A37B, which upgrades visual perception, cross‑modal alignment and complex reasoning, achieving top scores on image, video, and mathematical reasoning benchmarks while detailing its architecture, three‑stage pre‑training, post‑training strategies, and future multimodal agent plans.
This article breaks down the end‑to‑end engineering pipeline that converts a knowledge source such as a URL or PDF into a narrated PPT‑style video, detailing six core stages—from knowledge extraction and script generation to image creation, voice synthesis, and final video stitching—while highlighting practical model choices, prompt design, and stability tricks.
Gaode transforms its map app into a dynamic, AI‑driven “living map” by fine‑tuning the large Spacetime‑GR model through embedding‑based and generative ranking SFT, DPO alignment, and multimodal augmentation, achieving significant offline CTR‑AUC improvements and online CTR gains in POI recommendation.
VLM‑FO1 introduces a generate‑plus‑reference paradigm that replaces coordinate generation with region token referencing, adding plug‑in modules such as a proposal generator, a hybrid fine‑grained encoder, and a region‑language connector to give any pretrained visual language model accurate, fine‑grained perception while preserving its original capabilities.
This article examines the evolution from single‑frame video analysis to multimodal large models, detailing their architecture, optimization techniques, experimental validation on edge devices, and practical scenarios, while highlighting current limitations and future directions for AI‑driven video understanding.
LongCat-Flash-Omni, the latest open‑source model from Meituan, combines a 560 billion‑parameter architecture, efficient multimodal perception and speech reconstruction modules, and a progressive training strategy to deliver real‑time audio‑video interaction and state‑of‑the‑art performance across text, image, audio, and video tasks.
Artificial Analysis’s Q3 2025 AI report reveals a rapidly accelerating industry across the entire stack, with US and Chinese labs neck‑and‑neck, fierce competition among OpenAI, Google, Anthropic, xAI, DeepSeek and Alibaba, cost‑efficient models, booming multimodal agents, and a hardware race led by NVIDIA’s Blackwell accelerators.
This article explores the rise of AI agents—from their definition as intelligent digital assistants powered by large language models to their evolution through planning, memory, and tool use—highlighting real‑world applications, core technical mechanisms, code implementations, and future trends such as autonomy, multimodal fusion, standardization, and safety considerations.
This article explores how the explosion of unstructured data exposes the limits of traditional OCR and shows how emerging multimodal large language models provide end‑to‑end document understanding, reduce pipeline complexity, cut training costs, enable hybrid retrieval‑augmented generation, and drive real‑world industry deployments.
Learn how LangChain enables multimodal interactions by preparing inputs, constructing prompts, invoking models like GPT‑4o, and processing responses, with a complete example that demonstrates image‑question answering, code walkthrough, environment setup, and key considerations for API keys and image URLs.
Enterprise AI agents are moving beyond passive chatbots to become digital employees that perceive, reason, plan, learn, and act, and this article dissects a seven‑module agent platform architecture, its advantages, real‑world use cases, and future directions for scaling AI productivity across businesses.
Introducing Ming‑Flash‑Omni‑Preview, a 103‑billion‑parameter open‑source multimodal model built on a sparse MoE architecture that delivers state‑of‑the‑art performance in controllable image generation, streaming video understanding, and context‑aware speech recognition, surpassing prior models on GenEval and GEdit benchmarks.
Image GenerationSpeech Recognitionlarge language model
Google Research, X, and Cloud teams introduced Earth AI, a interoperable GeoAI model family that fuses image, population, and environmental data via a Gemini‑driven reasoning Agent, achieving state‑of‑the‑art performance and a 64% reasoning boost over Gemini 2.5 Pro while enabling non‑experts to run real‑time cross‑domain analyses.
Researchers present LucaOne, a Transformer‑based foundation model that unifies DNA/RNA and protein sequences using a 39‑token vocabulary, rotary positional encoding, and molecule‑type embeddings, and demonstrate through extensive multi‑task benchmarks that it outperforms domain‑specific models across seven biological tasks.
This week’s AI landscape saw Karpathy’s NanoChat open‑sourcing a 8‑K‑line ChatGPT replica, Ant Group unveiling a trillion‑parameter Ring‑1T model, Alibaba releasing the 4B/8B Qwen3‑VL visual language models that outperform Gemini 2.5 Flash Lite and GPT‑5 Nano, Google launching Veo 3.1 for high‑fidelity video generation, and Anthropic announcing Claude Haiku 4.5, a faster and cheaper LLM that excels on SWE‑bench benchmarks.
The article details Alibaba Cloud's MaxCompute transformation into an AI‑native data warehouse, highlighting its serverless elasticity, multimodal data management, unified model lifecycle, AI Function integration, and new distributed Python engine that together address the bursty, high‑complexity data and compute challenges of the generative AI era.
This article reviews Trading‑R1, an open‑source LLM inference framework that integrates multimodal financial data, three‑stage supervised‑fine‑tuning and reinforcement learning to generate structured investment arguments and risk‑adjusted trade decisions, achieving superior Sharpe ratio and drawdown performance on real‑world stock and ETF tests.
This article reviews Kuaishou's two‑year exploration of multimodal large‑model techniques for advertising, outlining challenges in content‑domain ad estimation, the COPE unified product representation framework, and the LEARN LLM knowledge‑transfer approach that together improve ad system performance.
This article reviews Kuaishou's two‑year exploration of large‑model techniques in advertising, outlining challenges in content‑domain ad estimation, introducing the COPE unified content representation framework and the LEARN LLM knowledge‑transfer approach, and showing how these innovations delivered tangible business gains.
The article examines the evolution from intelligent, assistant‑style testing to fully autonomous, LLM‑driven test agents, outlining four core capabilities, real‑world implementations across unit, API, and UI layers, and the technical pillars that enable self‑learning, self‑healing, and multi‑modal testing.
This article reviews Kuaishou's two‑year exploration of multimodal large‑model techniques for advertising, detailing challenges of fragmented user behavior, the COPE unified product representation framework, and the LEARN LLM knowledge‑transfer approach that together delivered measurable business gains.
This article presents Crab, a unified audio‑visual scene understanding model that leverages a novel display‑cooperation learning paradigm, introduces the AV‑UIE dataset with explicit reasoning steps, and demonstrates superior performance across temporal, spatial, pixel‑level, and spatio‑temporal tasks through extensive experiments and ablations.
This article reviews Kuaishou’s two‑year exploration of large‑model techniques in advertising, detailing the content‑domain estimation challenges, how multimodal and LLM approaches improve full‑domain behavior utilization and external knowledge integration, and introducing the COPE product‑content representation framework and the LEARN LLM knowledge‑transfer system.
OneCAT introduces a unified decoder‑only transformer that eliminates separate visual encoders, employs a modality‑specific MoE, integrates multi‑scale visual generation, and achieves state‑of‑the‑art performance and efficiency across multimodal understanding, text‑to‑image synthesis, and image editing tasks.
This article outlines Kuaishou's two‑year exploration of large‑model techniques in advertising, detailing challenges of sparse cross‑domain data, the COPE unified product representation framework, and the LEARN LLM knowledge‑transfer approach that together improve ad system effectiveness.
Over the past two years, Kuaishou has leveraged multimodal large‑model techniques to overcome sparse advertising data, integrating full‑domain user behavior and external knowledge via the COPE unified product representation framework and the LEARN LLM knowledge‑transfer system, achieving measurable business gains.
The article explains the concept of AI‑native applications, distinguishes them from AI‑plugin extensions, outlines their core principles such as model‑first design, data flywheel, event‑driven agents, multimodal semantics, continuous learning, and provides a seven‑step practical guide with code examples for building an AI‑native app.
This article presents concise English summaries of four recent arXiv papers that explore AI-driven trading frameworks, dual‑view risk‑relation identification from 10‑K filings, multimodal language models for financial forecasting, and credit‑spread prediction enhanced by non‑financial data, highlighting their methods, datasets, and performance results.
Manzano introduces a hybrid vision tokenizer and a three‑stage training recipe that let a 3‑billion‑parameter multimodal LLM achieve state‑of‑the‑art results on both image‑understanding benchmarks and text‑to‑image generation, while scaling smoothly to larger sizes and minimizing task conflict.
The article examines Mistral AI's rapid rise—from its Paris founding and record‑breaking seed round to ASML's €1.3 billion C‑round stake and Apple acquisition rumors—detailing its lightweight and multimodal models, open‑source strategy, product ecosystem, and the plagiarism and geopolitical debates that shape its valuation.
This article provides a detailed side‑by‑side analysis of ByteDance's Coze and Tencent's Yuanqi, examining their features, performance, ecosystem integration, free‑tier limits, target users, and future prospects to help developers and enterprises choose the platform that best fits their needs.
During a pre-conference interview, AI expert Wang Mingzhong details how multimodal AI dressing, video résumé creation, short‑video templates, and interactive digital‑human live streams are technically realized for 58 Home Services, highlighting model training, workflow optimization, and future fusion of template‑based and agent‑driven video generation.
Apple recently open‑sourced its FastVLM and MobileCLIP2 models, showcasing a multimodal vision‑language system that runs up to 85 times faster than comparable models, enabling real‑time AI on iPhones and other edge devices while illustrating Apple’s broader “B‑plan” of on‑device small‑model AI strategy.
This article explores how the LAS team's AI‑driven data lake solution, built on Daft for lake computing and Lance for lake storage, tackles the emerging challenges of multimodal data handling, offering faster I/O, heterogeneous CPU‑GPU scheduling, and seamless integration for AI workloads.
The article explains how to fine‑tune large multimodal models by focusing on the projection layer, optionally using LORA for language‑model adaptation, and highlights data alignment, common applications, and the added difficulty of modality alignment for interview preparation.
The article analyzes why enterprise‑wide large‑model AI projects face steep hurdles, outlining required human capabilities, historical labor shifts, current hot technologies such as RAG, Agent, CoT and multimodal, their limits, a three‑stage implementation roadmap, typical case pitfalls, and the key success factors for sustainable digital transformation.
This article introduces the rapidly popular Gemini CLI, compares it with Claude Code, explains its core features, demonstrates coding, multimodal, and MCP use cases, and details the author’s Easy LLM CLI fork that enables custom model integration, flexible configuration, and direct code embedding for developers.
This resource guide presents a curated list of cutting‑edge topics—including multimodal GraphRAG, knowledge‑graph‑driven large‑model applications in finance, traditional Chinese medicine, automotive manufacturing, and knowledge‑management trends—offering insights into AI‑powered knowledge services, and invites readers to scan the QR code to download the full e‑book.
This article explains how embeddings transform discrete items such as text, images, or user actions into continuous vectors, walks through the step‑by‑step workflow—from tokenization to normalization—highlights core properties, compares popular models, and showcases practical use cases in e‑commerce intent filtering and medical image retrieval, all backed by concrete examples and code.
The Kwai Keye team presents Thyme, a novel multimodal reasoning framework that lets large language models generate and safely execute Python code for image manipulation and complex calculations, achieving significant performance gains over existing vision‑language models across perception, reasoning, and hallucination‑reduction benchmarks.
Data‑Juicer is an open‑source, one‑stop multimodal data processing system that provides fine‑grained operators, scalable pipelines, and ready‑made recipes to deliver high‑quality, diverse, and model‑friendly data for large language model pre‑training, fine‑tuning, and multimodal applications.
The 2025 H1 AI Core Achievements and Trends report reveals how agents are reshaping productivity, models are gaining inference power and becoming smaller, reinforcement learning is overtaking pre‑training, and industry competition is intensifying, with China and the US narrowing their technology gap.
The Q2 2025 State of AI report analyzes Chinese AI labs’ rapid progress across language models, open‑source weights, and multimodal generation, showing a shrinking performance gap with US leaders, detailed benchmark scores, ecosystem classifications, and emerging competitive dynamics.
This article presents HiddenDetect, a training‑free method that leverages refusal‑semantic vectors and layer‑wise activation analysis to detect jailbreak attempts in multimodal large language models, revealing distinct safety signals across text and image modalities and demonstrating strong performance on several LVLM benchmarks.
VeOmni, ByteDance’s open‑source unified multimodal training framework, tackles fragmented training pipelines by integrating LoRA fine‑tuning, FSDP, Ulysses, and Expert Parallel, delivering up to 40% higher throughput, up to 55% memory savings, and streamlined one‑click deployment for LLM, VLM, and video models.
The report analyzes first‑half‑2025 AI breakthroughs, covering the rise of general‑purpose agents, rapid inference improvements, small‑model proliferation, reinforcement‑learning compute dominance, evolving transformer architectures, and shifting industry dynamics, offering actionable insights for researchers, product leaders, and decision‑makers.
Lumina-mGPT 2.0 is a decoder‑only, zero‑shot trained autoregressive image model that rivals diffusion systems like DALL·E 3 in quality while offering unified multimodal tokenization, flexible multi‑task generation, and several inference‑speed tricks, yet it still faces licensing, scaling and sampling‑time challenges.
This article presents a curated list of cutting‑edge resources covering multimodal GraphRAG, knowledge‑graph‑driven large‑model applications in finance, healthcare, automotive, and more, offering insights into the evolving synergy between AI and knowledge graphs.
This resource compiles a curated list of cutting‑edge studies and frameworks that combine multimodal GraphRAG, knowledge‑graph techniques, and large‑model AI across sectors such as finance, healthcare, and manufacturing, guiding readers through emerging paradigms and future directions.
This presentation explores the evolution of AI agents in e‑commerce content creation, detailing the transition from text‑only industrial production (1.0) to multimodal image and video generation (2.0) and finally to quality‑driven optimization and decision‑making (3.0), highlighting technical architectures, challenges, and future directions.
The article reviews Grok 4’s flashy launch and claimed first‑principles advantage, then presents benchmark results—showing strong reasoning, multimodal and agent scores but disappointing coding performance versus DeepSeek‑R1—concluding that the model’s real‑world capabilities fall short of its hype.
This article explains a system that combines large language models with multimodal vector retrieval to automatically understand user intent and instantly recommend the most relevant icons, detailing the workflow, semantic vectorization, offline indexing, online inference, and evaluation methods.
Elon Musk unveiled Grok‑4, a subscription‑based AI reasoning model that claims near‑human performance on elite exams, showcases unprecedented benchmark scores, multimodal understanding, voice synthesis, and a roadmap of upcoming coding and video generation models, while introducing a $30/month and $300/month tier.
Kuaishou has had eight cutting‑edge papers accepted at the International Conference on Machine Learning 2025, covering breakthroughs in multimodal emotion modeling, monotonic probability learning, causal effect generalization, cascade ranking, multimodal LLM alignment, ultra‑low‑rate image compression, and visual autoregressive super‑resolution, with links to each work and accompanying code repositories.
This article presents a comprehensive overview of cutting‑edge research on integrating large language models with knowledge graphs, covering multimodal GraphRAG, financial AI solutions, traditional Chinese medicine decision support, and industry‑specific knowledge services, guiding readers through emerging paradigms and practical implementations.
In a QECon 2025 talk, Ant Group’s AI team detailed a comprehensive multimodal evaluation framework that leverages large‑model metrics, custom pipelines, and benchmark datasets to assess image generation, speech recognition, and video quality, while also contributing to industry standards and academic research.
In an interview with Vivo AI engineer Liang Tianan, the article explores the challenges of post‑Q&A recommendation, the integration of large language models into recall, ranking and evaluation pipelines, and the engineering trade‑offs required to deliver high‑quality, diverse suggestions on mobile devices.
In an interview at the DA Digital Intelligence Conference, ByteDance AI specialist Cai Conghuai explains how large language models, combined with techniques like SFT, DPO, and RAG, are reshaping Douyin's user‑experience signal detection, root‑cause analysis, and evaluation, while outlining future AI‑agent breakthroughs.
This article outlines the design of a scientific, quantifiable, multi‑dimensional evaluation system for the DataV‑Note intelligent analysis platform, addressing the lack of unified standards and accuracy challenges in AI‑driven data reporting, and proposes concrete metrics, model architecture, and future automation plans.
This article chronicles the rapid development of large language models from the 2017 Transformer breakthrough through the rise of BERT, GPT‑3, multimodal models, alignment techniques like RLHF, and finally the cost‑efficient DeepSeek‑R1 in 2025, highlighting key innovations, scaling trends, and real‑world impacts.
OpenAI introduced the O3‑Pro multimodal deep‑reasoning model with an 80% price cut for O3, detailed its training via large‑scale reinforcement learning, compared its capabilities and costs against GPT‑4o, GPT‑4.1 and O3‑Pro, listed its core specs, limitations, access methods, and presented benchmark tests that highlight both strengths and weaknesses.
IDC’s 2025 China foundational large‑model evaluation crowns Baidu’s Wenxin as the top performer, scoring perfect marks in seven of eight criteria and highlighting its superior multimodal, dialogue, and ecosystem capabilities among twelve leading models.
Kuaishou presented twelve peer‑reviewed papers at CVPR 2025 covering video quality assessment, large‑scale video datasets, dynamic 3D avatar reconstruction, 4D scene simulation, controllable video generation, scaling laws for diffusion transformers, multimodal foundations, and more, highlighting the company's leading research in computer vision and AI.
Quwan Technology presents its Kaitian social large model, designed for personalized, emotionally rich, multimodal AI interactions, detailing its scene‑specific goals, CPT+SFT+RLHF training pipeline, data desensitization, LoRA fine‑tuning, evaluation methods, pruning, latency trade‑offs, safety mechanisms, and future feedback loops.
Google AI Edge Gallery lets a 2 GB RAM Android phone run sophisticated multimodal AI models completely offline, offering image understanding, text generation, and conversational capabilities without any network connection, and provides a quick three‑step setup to start experimenting with on‑device AI.
Kuaishou’s foundational large-model team has secured seven papers at ACL 2025, spanning alignment bias in training, safety defenses during inference, decoding strategies, fine-grained video-temporal understanding, reward fairness in RLHF, multimodal captioning benchmarks, and methods to curb hallucinations in vision-language models.
This article provides a comprehensive technical analysis of LLM context, covering its definition, types, tokenization, window‑size evolution, diminishing returns, management techniques such as RAG, CoT, memory‑as‑a‑service, and future challenges like multimodal fusion, privacy, and autonomous agent memory.
This roundup highlights ten recent AI developments, from Apple's Matrix3D model that creates 3D scenes from photos, to Qwen's Deep Research assistant, Tencent's CodeBuddy 3.0, ByteDance's Seed1.5‑VL, Step Star's open‑source Step1X‑3D, Google's iOS icon refresh, Apple's eye‑tracking scrolling test, Chrome's upcoming Gemini AI assistant, Shanghai's AI Identity Ecosystem Alliance, and Kuaishou's Keling AI 2.0 topping the global video‑generation leaderboard.
The 2025 Q1 AI report from Artificial Analysis highlights six major trends—including a thousand‑fold drop in inference cost, the rise of MoE models, the growing parity of Chinese open‑source labs, the emergence of autonomous AI agents, native multimodal capabilities, and the trade‑off between performance, cost, and context windows—painting a picture of a rapidly evolving, increasingly competitive AI ecosystem.
At Baidu AI Day in Beijing, the company unveiled the Wenxin 4.5 Turbo and X1 Turbo models, detailing multimodal training breakthroughs, self‑feedback loops, enhanced reasoning and tool‑calling, while the China Academy of Information and Communications Technology awarded X1 Turbo the highest "4+" rating across 24 capability tests, highlighting its leading position in domestic large‑model performance.
This article examines the evolution of large model technologies—including Retrieval‑Augmented Generation, AI agents, and multimodal models—detailing their technical foundations, practical challenges, industry applications, and future development trends, offering a comprehensive perspective for AI practitioners and researchers.
The article introduces FineVD, the first large‑scale multi‑dimensional UGC video quality dataset, and presents FineVQ, a unified model that predicts quality scores, attributes, and distortion types across six dimensions, achieving state‑of‑the‑art performance on multiple benchmarks and cross‑dataset evaluations.
This article introduces OpenAI’s CLIP multimodal model, explains its architecture and contrastive training, details hardware and installation steps, and demonstrates a hands‑on zero‑shot image classification workflow that achieves 97% confidence on a cat image without any task‑specific fine‑tuning.
This article surveys the rapid emergence of AI agents, outlining their projected 2025 breakthrough, market momentum, key frameworks such as Manus and MCP, the four core abilities of perception, planning, tool use, and memory, and the evolving landscape of multimodal and autonomous AI systems.
This talk explores the technical challenges of applying large language models and knowledge graphs in finance, discusses solutions such as RAG enhancements, graph‑guided retrieval, multimodal extensions, and presents future research directions including multimodal graph integration, agentic systems, and decision‑making applications.
The article presents the Universal Recommendation Model (URM), a large‑language‑model‑based recall framework that integrates world knowledge and e‑commerce expertise through knowledge injection and prompt‑driven alignment, achieving significant offline recall gains and a 3.1% increase in ad consumption while meeting high‑QPS, low‑latency production constraints.
An award‑winning joint project by Shanghai Jiao Tong University and Ant Group unveils a multimodal foundation model and advanced detection techniques that dramatically improve AI‑driven content risk governance across massive online services.
AI agents are rapidly evolving toward a 2025 breakthrough in perception, autonomous planning, tool use and memory, driven by multimodal models, neural‑symbolic reasoning and embodied intelligence, with $27 billion investment forecasts, exemplified by general‑purpose agents like Manus and emerging applications in code generation, research, healthcare, and risk analysis.
This article introduces Spring AI's multimodal capabilities, explains the Message API for handling text, image, audio, and video inputs, and provides step‑by‑step Spring Boot examples for image analysis, ID card extraction, and structured JSON output of car‑color counts.
The Qwen3 series introduces eight open‑source large language models ranging from 0.6B to 235B parameters, combines dense and Mixture‑of‑Experts architectures, supports multimodal input, offers mixed inference modes, and demonstrates benchmark superiority over leading models such as OpenAI o1 and Gemini 2.5 Pro.
This article chronicles the evolution of large language models from the 2017 Transformer breakthrough through BERT, GPT series, multimodal models, and recent cost‑efficient innovations like DeepSeek‑R1, highlighting key architectures, training methods, alignment techniques, and their transformative impact on AI applications.
This article examines how Retrieval‑Augmented Generation (RAG), AI agents, and multimodal large‑model techniques are reshaping AI‑industry integration, discusses their technical challenges and practical implementations, and outlines future development directions across algorithms, products, and domain‑specific applications.
This article compiles a ranked list of ten popular open-source OCR projects on GitHub, summarizing each tool’s key capabilities—such as multimodal text extraction, PDF linearization, layout analysis, and multilingual support—along with star counts and direct repository links for developers seeking ready-to-use OCR solutions.
This article examines how a front‑end team uses AI innovations—FastGPT visual platform, AI‑powered semantic search, and AI video synthesis—to rebuild business workflows, cut costs, and boost efficiency, highlighting architecture, technical highlights, and practical use cases.
The article demonstrates how ByteDance’s new Doubao‑1.5 multimodal model can be used to locate and count objects in images—such as sushi plates, street signs, and cartoon hats—by generating coordinates and overlaying visual annotations through a concise Python script.
This article provides an in‑depth analysis of DeepSeek’s Janus and Janus‑Pro models, explaining how decoupling visual encoding resolves the conflict between multimodal understanding and generation, detailing training stages, data scaling, architectural choices, and presenting extensive benchmark results that demonstrate significant performance gains.
Zero‑UI, powered by multimodal AI models like Google Gemini, is shifting design from screen‑based interfaces to natural voice, gesture, and environmental interactions, prompting a fundamental redesign of how devices understand user intent across smart homes, cars, and immersive experiences.
The China AI Industry Alliance unveiled its Q1 2025 Fangsheng benchmark, showing Baidu’s new multimodal models—Wenxin 4.5 leading basic abilities and Wenxin X1 excelling in reasoning—available for free on the Wenxin Yiyan platform, while Baidu pledges major 2025 investments in AI, data‑center and cloud infrastructure.
The article presents a series of technical breakthroughs by JD's advertising team that improve the quality and coverage of AI‑generated ad images through a trustworthy multimodal feedback network, introduce a large human‑annotated image dataset, and enhance creative ranking with offline multimodal representations and online architecture optimizations, ultimately achieving more precise and scalable ad personalization.
This report details a comprehensive set of optimizations for multimodal visual large‑model (VLM) inference—including image pre‑processing acceleration, TensorRT integration for the ViT module, CUDA‑Graph replay, token‑count reduction, prefix‑cache handling, and weight quantization—demonstrating up to three‑fold throughput gains while maintaining accuracy.
At the ICLR 2025 live session in Singapore, Ant Group showcased four cutting‑edge papers—CodePlan, Animate‑X, Group Position Embedding, and OmniKV—demonstrating advances in large‑language‑model reasoning, universal character animation, layout‑aware document understanding, and efficient long‑context inference.
AI researchDocument UnderstandingLarge Language Models
On April 2, Baidu released its Wenxin X1 large model on the Qianfan platform, offering enterprise users and developers a multimodal, deep‑thinking AI with superior math, coding, and reasoning scores, low token‑price API access, batch inference, one‑click distillation, and rapid RAG/Agent application building.
A recent study by LeCun, Xie, and collaborators shows that large‑scale visual self‑supervised learning (Web‑SSL) can match or surpass CLIP on diverse VQA tasks, even without any language supervision, by scaling model size and data volume.
Meta’s newly released Llama 4 models—Maverick with 4 020 billion total parameters and Scout with 1 090 billion—feature a 128‑expert MoE, 10 million‑token context, native multimodal fusion, and FP8 training, delivering benchmark‑leading performance that outpaces GPT‑4o, Gemini 2.0 Flash and DeepSeek v3, while being openly available on Hugging Face and GitHub.
The article analyses Google’s Gemini 2.5 Pro as a decisive shift toward a “Reasoning Model”, detailing its architectural focus on inference, benchmark breakthroughs such as Humanity’s Last Exam and GPQA Diamond, long‑context capability, multimodal strengths, Vibe‑coding experience, and the roadmap for future Gemini models.
This article reviews AutoGLM, the first "think‑while‑doing" AI agent released by Zhipu AI, detailing its core capabilities, full‑stack architecture, user experience, identified limitations, and the outcomes of three hands‑on tests using both the client application and a Chrome extension.
VBench-2.0 expands the original VBench suite by introducing six fine‑grained dimensions—Human Fidelity, Controllability, Creativity, Physics, Commonsense, and more—to evaluate not only the visual quality of generated videos but also their intrinsic faithfulness to physical laws, common sense, and narrative coherence, providing open‑source tools, prompts, and human‑aligned metrics for the research community.