Not bad for limited hardware ...

Unitree Robotics humanoid robots dance during the opening day of its Asia's first embodied intelligence experience store in Shanghai on May 31, 2026. Jade GAO/Getty Images China's government issued a ...

A new 3D image projection system marks a major step toward overcoming a longstanding problem for holographic technology.

AI medical imaging market is projected to exceed $20B by 2035. Generative models address class imbalances in medical imaging ...

好家伙，这次不是模型圈自嗨。 一个训练成本约1500美元、参数量约1B、从零开始预训练的小模型，把 HRM推到了下一代推理架构讨论的中心。 HuggingFace联合创始人兼CEO Clem Delangue亲自转发推荐。 图灵奖得主Yoshua ...

好家伙，这次不是模型圈自嗨。 一个训练成本约1500美元、参数量约1B、从零开始预训练的小模型，把HRM推到了下一代推理架构讨论的中心。 HuggingFace联合创始人兼CEO Clem Delangue亲自转发推荐。 图灵奖得主Yoshua Bengio作为共同作者参与的新论文，也走向了同一条latent recursive reasoning路线。 更反常的是，它不是蒸馏，不是微调，也不是在 ...

With LLMs increasingly working multimodally, there are exciting developments for more performance and leaner sizes.

With LLMs increasingly working multimodally, there are exciting developments for more performance and leaner sizes.

LCLMs compress LLM context before decode — 8.8x faster at 16x compression, beating every KV cache method tested. Open-sourced by NYU and Columbia.

过去一年，开源模型的发布节奏已经快到让人麻木。每次发布，伴随的永远是一组跑分、一张能力雷达图，以及几个“超越某某”的结论。

Scalable data fusion for generating disaggregate activity schedules ...

2K 图像 210ms 解码，4K 细节直接生成，传统「解码 + 超分」流水线可能要被重写了。 文生图模型现已内卷至 4K、超写实、电影级光影水准，但一张 AI 图像要想像 GPT-Image2 ...