Tag
NVIDIA 发布 Cosmos3-H-Dreams，这是首个基于临床腹腔镜胆囊切除手术记录训练、支持实时运动学控制的交互式生成式手术模拟器，可在单块 NVIDIA RTX PRO 6000 GPU 上流畅运行。
SolarRenderer 将任意引擎的无纹理结构通道结合参考首帧，生成风格化完整视频，基于 MiniMax-H3 构建，通过控制 LoRA 与少步蒸馏 LoRA 实现实时逐帧结构一致的渲染，无需按游戏训练。
Matrix-Game 4.0 将游戏逻辑与视觉合成分离，通过可执行引擎与自回归视频渲染器实现低延迟的实时可玩游戏世界，在单张 RTX5090 上达到 480p、60 FPS 及 112ms 按键到显示延迟，并开源代码、模型权重与低多边形游戏资产。
SteerTTS 是一个 23M 参数的可数值调控英文 TTS 模型，通过 z-score 控制语速、音高、响度等 12 项参数，可在 CPU 上实时运行，并附带 2825 个内置语音与实验性零样本克隆。
论文提出 RealtimeBench，包含六类游戏共180个任务，在推理期间游戏持续运行来评估 LLM 智能体的实时决策能力，并提出 TORSI 双时间尺度自我改进方法，将 Jev 模型平均分从 13.51 提升至 24.19。
Grok Voice 3.0 发布，被称为全球最智能的语音到语音模型，响应时间低于1秒，面向客服通话场景且成本低于其他前沿模型。
A 1.4M-parameter U-Net student model, distilled from FLUX.2 klein, achieves real-time 30-40 FPS neural weather restyling (snow, wet, night) inside a Minecraft Fabric mod on a budget GTX 1650, using a PatchGAN fine-tune to fix washed-out outputs from pure pixel losses.
作者重新评估Jev类模型的价值，认为随着模型足够快和便宜，原本靠if-else拼接的业务逻辑、实时语音交互、计算机操作和机器人决策都可以交给System 1模型处理，形成SOTA后台深思考模型与System 1实时模型分工的新范式。
作者将六个流行的决策模型（jev 1.13、Kev 4B、Clef、Clef Flash、GPT-6 Luna 和 Laya）放在《吃豆人》游戏中进行实时对决，jev 1.13 以平均 2,750 分领跑排行榜。团队已公开排行榜和开源仓库，任何人提交自己的模型即可参赛。
WorldCast 提出了一种分布式多人世界模型，每个玩家本地运行视频生成器与状态模型，通过交换玩家状态和共享场景状态实现一致、可扩展的实时多人生成，在 Counter-Strike 2 上验证了其一致性与可扩展性。
A solo researcher shares Part 2 of a local AI world model: a ~960M-parameter transformer using diffusion forcing and text cross-attention that runs at 30-60 FPS on an RTX 5090, accepts live WASD actions, and supports mid-rollout text prompt switching (e.g., 'add a pond to the desert').
Clef Flash, a 9B model running at Q4 quantization on an RTX 5080, plays Google Snake in real time (within the game's 135ms turn limit) using only simple vision instructions, with no training or game-state hacking required.
dotwave.ai shares technical write-ups showing their inference engine can serve 56 concurrent real-time sessions on a single H100 (vs 1 for NVIDIA's reference stack) by keeping the model resident on the GPU and batching all sessions per tick, cutting per-conversation GPU cost by ~98% with zero missed 160ms audio deadlines.
Microsoft AI 发布首个流式转录模型 MAI-Transcribe-2-Streaming（在 Artificial Analysis 上排名准确度第一，60 种语言、100ms 部分结果延迟），同时推出 MAI-Voice-2.1 与 MAI-Voice-2.1-Flash 语音合成模型，面向低延迟对话式语音智能体场景。
Maxime Labonne shares a demo of an AI agent playing Super Smash Bros and beating a Level 9 CPU Fox two stocks on Final Destination in real time through an API, highlighting advances in real-time agentic gaming.
LangChain hosted a webinar discussing Jev, a tool that enables real-time inference for tasks like content filtering and guardrails, featuring insights from Sydney Runkle and others.
AI capabilities for expression recognition, object detection, and finger counting have achieved less than 1 second latency, highlighting notable real-time performance gains.
The paper introduces ZENDAYA, a closed-form bandwidth dial for streaming multimodal decoders that dynamically adjusts input reading to improve real-time text generation performance. It demonstrates that reducing input consumption can enhance quality and efficiency in streaming settings across video and audio benchmarks.
The Atlas model now runs in real-time after intense inference optimization, making world generation and navigation more interactive on both B200 and MI355X hardware.
ReactVAU introduces a slow-fast decoupled framework for real-time streaming video anomaly understanding, leveraging a fast detection module, persistent anomaly-aware memory, and on-demand slow reasoning to enhance efficiency and performance.