Tagged articles
real-time interaction
28 articles · Page 1 of 1
NVIDIA and UT-Austin introduce Realtime-Bench, a benchmark testing 12 frontier models on real-time interactive tasks where environments evolve continuously; even the best model GPT-6 Astra achieves only 44% success, revealing that increased reasoning can hurt performance due to time constraints, and reducing latency alone is insufficient.
AI-evaluationGPT-6NVIDIA
0 likes · 11 min read

openJiuwen WorkSwarm introduces a full-duplex multimodal AI agent platform where a real-time model handles live audio/video interaction while a Core Agent executes complex tasks in the background, allowing users to interrupt, reprioritize, and continue chatting without waiting for tool calls to finish.
Ascend NPUbackground task executionfull-duplex
0 likes · 11 min read

openJiuwen WorkSwarm introduces full-duplex multimodal interaction where a real-time model handles audio/video while a Core Agent runs background tasks, allowing users to interrupt, delegate work, and manage a persistent task queue across Windows, Mac, and HarmonyOS with support for JoyAI, Qwen Omni, and vLLM-Omni on Ascend NPU.
AI agentAscend NPUHuawei
0 likes · 9 min read

Alibaba's Qwen3.8-Omni-Flash slashes API costs by 98% for audio and 93% for audio-video, offers 1M context, beats Qwen3.5-Omni-Plus by 25% on benchmarks, surpasses Gemini 3.8 Flash in audio understanding, and introduces real-time spatial audio localization for robotics.
AlibabaGemini 3.8 FlashQwen3.8-Omni-Flash
0 likes · 10 min read

VoiceMem introduces a brain-inspired dual-memory system for real-time voice AI, separating factual and emotional memory to achieve 134ms retrieval latency and enable long-term personalized companionship.
Emotional AILLM memoryMemory Systems
0 likes · 11 min read

Video Rebirth builds Olympus, a video‑native, real‑time interactive world model that emphasizes physical, audio‑visual, and long‑term memory consistency, leveraging its BACH video generation engine, multi‑card inference optimizations, and AMD MI350 hardware to bridge digital and physical AI domains.
AI consistencyAMD MI350BACH engine
0 likes · 11 min read

MoWorld introduces a Flash World Model that runs on domestic Ascend NPU, delivering cinematic‑grade quality at up to 50 FPS real‑time interaction while cutting inference cost by 70%, showcasing a breakthrough in low‑cost, high‑frame‑rate world‑model deployment.
14B MoEFlash World ModelHigh Frame Rate
0 likes · 12 min read

A two‑hour Douyin live broadcast demonstrated an AI‑driven interactive game where the AI acted as scriptwriter, host and scheduler, handling multimodal inputs, real‑time state management and fault‑tolerant runtime, achieving 935k total exposures and 29k peak concurrent viewers while redefining live‑stream participation.
AI agentAgent RuntimeComplexity Engineering
0 likes · 17 min read

JD’s open‑source JoyAI‑VL‑Interaction transforms large‑model AI from static question‑answering to continuous, on‑scene observation, proactive judgment, and real‑time response, offering agent delegation and achieving up to 87.9% win rate against leading video assistants in live benchmarks.
AI assistantVision-Language Modelbenchmark
0 likes · 9 min read

Li Mu (沐神) returns to Bilibili after a year to showcase Higgs Avatar v1, a fully AI‑generated real‑time digital human that can listen, speak, lip‑sync and display facial expressions, with performance metrics showing 16 ms per frame on a single H100 GPU and potential applications ranging from customer service to training, while also raising ethical considerations about identity and trust.
AI AvatarBoson AIDigital Human
0 likes · 7 min read

The article clarifies the dual meaning of “end‑to‑end” in speech AI—product simplicity and engineering unification—then outlines six emerging trends, from real‑time conversational latency to multilingual robustness, token‑based audio pipelines, voice‑specific security, edge privacy, and the growing importance of data quality and reproducibility.
Edge ComputingEnd-to-EndSpeech AI
0 likes · 8 min read

Alibaba’s ATH team unveiled HappyOyster, a real‑time world‑model platform that lets users generate and explore interactive 3D environments from a single sentence or image, offering two modes—Wander for exploration and Direct for creation—while detailing its streaming architecture, multimodal foundation, competitive advantages, use cases, and current limitations.
AI videoGenerative AIWorld Model
0 likes · 11 min read

AURA introduces an always‑on video LLM that processes streams frame‑by‑frame, decides when to stay silent or answer, uses a dual sliding‑window context and a Silent‑Speech Balanced Loss, achieves state‑of‑the‑art scores on StreamingBench, OVO‑Bench and OmniMMI, and runs at 2 FPS with ~312 ms end‑to‑end latency on two 80G GPUs.
AURASilent-Speech LossStreaming Video Understanding
0 likes · 15 min read


21CTONov 4, 2025 · Artificial Intelligence
LongCat-Flash-Omni, an open‑source 560 billion‑parameter multimodal model, combines efficient Shortcut‑Connected MoE architecture with advanced perception and speech modules to deliver low‑latency real‑time audio‑video interaction and state‑of‑the‑art performance across text, image, video, and audio tasks.
audio-visual processingefficient inferencelarge language model
0 likes · 10 min read

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.
AIbenchmarklarge language model
0 likes · 9 min read

This article explains how a cloud testing platform leverages WebRTC to achieve sub‑200 ms bidirectional video transmission, enabling ultra‑low‑latency screen casting and remote camera feed replacement for mobile devices, and details the architecture, optimizations, performance gains, and future enhancements.
Mobile AutomationWebRTCcloud testing
0 likes · 20 min read

II-Agent is an open‑source, multi‑domain AI agent framework that leverages powerful large language models, a rich toolset, planning‑and‑reflection mechanisms, and advanced context management to enable autonomous task execution, real‑time interaction, and seamless integration across development, data analysis, and enterprise workflows.
AI agentContext Managementautomation
0 likes · 21 min read

The AG-UI (Agent User Interaction) protocol standardizes communication between backend AI agents and front‑end interfaces using a single JSON event stream, addressing real‑time streaming, tool orchestration, shared state, concurrency, security, and framework fragmentation to enable seamless human‑agent collaboration.
AG-UIAI agentsBackend
0 likes · 8 min read

WonderTurbo introduces a real‑time 3D scene generation pipeline that accelerates both geometry and appearance modeling to under a second per view, using StepSplat, QuickDepth, and FastPaint modules, achieving up to 15× speedup while maintaining high visual quality.
3D generationComputer VisionDepth Completion
0 likes · 16 min read

This article explains the design, implementation, and underlying Agent‑Oriented‑Programming framework of NetEase Fuxi’s multimodal real‑time voice AI teammate for the mobile game ‘Naraka: Bladepoint’, highlighting its capabilities such as autonomous navigation, combat assistance, natural dialogue, teaching, and broader applications of voice technology in games.
Naraka Bladepointagent-oriented programminggame AI
0 likes · 12 min read

To solve duplicated functionality, legacy code, and scalability limits in Bilibili’s live‑streaming interaction services, the team created a unified Interaction Center that abstracts RTC, consolidates session, link, UI, scoring and role management, introduces a shared state machine and tracing, and evolves through phased, extensible architecture for higher performance and maintainability.
Performance MonitoringRTClive streaming
0 likes · 22 min read

Bilibili’s live‑streaming tech team created an open interactive platform—spurred by the 600,000‑viewer success of Xiu Gou Nightclub—that supports hang‑up, host‑enhanced, and tool‑assisted streams, provides SDKs, APIs, data‑compliant authentication, tackles latency and rendering challenges, and now explores advertising, sponsorship and game‑promotion models to sustain its ecosystem.
SDK DevelopmentWebSocket communicationdata compliance
0 likes · 13 min read

This article provides a comprehensive overview of Volcano Engine's virtual digital human platform, detailing its definition, AI‑driven and human‑driven classifications, 2D and 3D technical architectures, multi‑modal perception, interaction capabilities, application scenarios, and future development directions.
2D avatar3D avatarComputer Vision
0 likes · 15 min read

The paper presents a cloud‑native architecture for AI‑driven digital‑human live‑streaming, detailing three‑layer asset, interaction, and media modules, real‑time script and Q&A scheduling, fault‑tolerant rendering and control services, and demonstrates how virtual anchors can deliver continuous, lifelike 24/7 e‑commerce streams.
AIDigital Humancloud rendering
0 likes · 21 min read

TWebLive, Tencent Cloud’s new web‑live interactive SDK, bundles TRTC, TIM and TCPlayer to let developers add push streaming, low‑latency WebRTC or CDN playback, and real‑time chat or bullet‑screen interaction with simple APIs, demo projects and open‑source code, replacing legacy Flash solutions.
JavaScriptSDKTencent Cloud
0 likes · 11 min read

Alibaba Entertainment’s 2019 Double‑11 "Cat Night" showcased a suite of cutting‑edge streaming technologies—including multi‑angle frame alignment, Dolby‑Atmos audio, low‑latency SRT transport, smart bitrate, edge‑cloud distribution, and a zero‑loss quality‑assurance system—that enabled a seamless 4K experience for viewers in over 200 countries.
Dolby Atmosedge-cloudglobal scale
0 likes · 9 min read

This article details the research and engineering behind Alibaba's AliPlayStudio, a video‑interactive platform that combines computer‑vision algorithms such as human parsing, gesture and pose detection, and controllable style transfer, all optimized for real‑time deployment on low‑power mobile and embedded devices.
gesture recognitionmobile AIpose estimation
0 likes · 17 min read

The article details Alibaba's technical innovations for the 2016 Double‑11 live event, covering two‑way audience interaction, time‑offset synchronization, massive real‑time like ranking, AR cross‑screen features, and the custom internet‑director console that together enabled seamless, high‑concurrency, multi‑platform engagement.
ARBackend Engineeringhigh concurrency
0 likes · 14 min read
