@KwaiKeye profile picture
KwaiKeye @KwaiKeye
Joined May 2026
  • Tweets 22
  • Following 12
  • Followers 101
3 Photos and videos
Pinned Tweet
KwaiKeye@KwaiKeye
May 26
🚀 Keye-VL-2.0-30B-A3B released — fully open. The first VLM to land DSA (DeepSeek Sparse Attention) in production — nearly lossless reasoning over 256K ultra-long context. 30B params. Fully open. 🤗 🧵👇
1
4
12
318
🎉The Kwai Keye-VL-2.0 Technical Report is now live. Keye-VL-2.0 is the first to adapt DeepSeek Sparse Attention (DSA) to GQA-based multimodal architectures, enabling lossless 256K context processing for long-video understanding. Paper: huggingface.co/papers/2606.1…

Paper page - Kwai Keye-VL-2.0 Technical Report

Join the discussion on this paper page

huggingface.co
2
47
🎆 Keye-VL-2.0实测案例：识别非遗炭花舞的视觉与情感双重高潮 面对一段记录传统炭花舞的视频，当提问“找到视频中的高光时刻”时，Keye-VL-2.0- 30B-A3B 没有陷入传统 AI “哪里最亮就截取哪里”的机械逻辑，而是基于视觉冲击力与情感共鸣，给出了极具审美深度的分段式高光判定： 👇🏻🧵
6:06
1
2
54
more replies
[05:33-05:46]情感共鸣的升华收尾： 模型判定依据： 表演结束后观众群情激昂，爆发热烈掌声与欢呼声。人群纷纷举起手机记录，共同沉浸于表演魅力中。这一段落标志着非遗技艺成功突破地域限制并获得大众共鸣，社会意义在此刻升华，是全片的情感高光与叙事收尾点。
1
1
33
现在体验👇🏻： Hugging Face： huggingface.co/Kwai-Keye/Key… GitHub：github.com/Kwai-Keye/Keye （致谢：本文案例演示的视频素材源自快手平台 @南翔）

Kwai-Keye/Keye-VL-2.0-30B-A3B · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
1
31
🎬 Keye-VL-2.0实测案例：精准拆解非遗茶器的制作时序 向 Keye-VL-2.0-30B-A3B 输入了一段制作陶杯的工艺流程视频。面对繁复的手工动作，模型展现出了手术刀般的帧级判断力，直接输出了一份带有精确时间戳的工艺全拆解 🧵👇
9:34
1
1
8
409
more replies
4.茶杯坯体制作与装饰：转轮手工拉坯塑形；精细修整厚度形状；底部贴方形印章款识并修整 5.配釉与施釉：按比例称取石英石长石等原料并加水搅拌制浆；坯体反复浸入釉浆后自然晾干 6.烧制与成品展示：装窑堆叠；烧柴升温至 1200℃；出窑清洗浸入陈茶水中氧化调整釉色；最终呈现开片、铁足等釉色特征
1
273
从挖土、煅烧、水飞到修坯、配釉、陈茶，全程极具专业壁垒的手工工序，模型不仅全部准确识别，更将每一个工艺动作与视频时间轴实现了毫秒级的完美咬合。 现在体验👇🏻： Hugging Face： huggingface.co/Kwai-Keye/Key… GitHub：github.com/Kwai-Keye/Keye （致谢：本文案例演示的视频素材源自快手平台 @山白）

Kwai-Keye/Keye-VL-2.0-30B-A3B · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
2
4
394
🚀 Keye-VL-2.0-30B-A3B released — fully open. The first VLM to land DSA (DeepSeek Sparse Attention) in production — nearly lossless reasoning over 256K ultra-long context. 30B params. Fully open. 🤗 🧵👇
1
4
12
318
▸ Tops video benchmarks at its scale ▸ Outperforms Qwen3-VL-235B on LongVideoBench (74.1) ▸ First Keye base model with built-in Agent: Code · Tool · Search
1
3
151
Find us on Hugging Face & GitHub 👇 Hugging Face： huggingface.co/Kwai-Keye/Key… GitHub：github.com/Kwai-Keye/Keye

Kwai-Keye/Keye-VL-2.0-30B-A3B · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
2
1
3
162
Load more