Tencent released Hy4 Flash, a multimodal MoE model (~220B–224B total parameters, 12B active) with a 1M-token context window. Combining hybrid attention, LatentMoE, and Engram lookup structures cut KV cache size to one-sixth of Hy4 while reaching over 150 tokens/s with 4-bit quantization and megakernels.
Chinese AI Labs
Also known as: Chinese Labs · China AI Labs
Plain English
Chinese AI labs include groups that build language models, systems that can read images and video, and tools that generate video from prompts or reference images. This overview follows selected releases from several of those teams.
Technical Overview
Alibaba’s Qwen work includes vision-language models, while DeepSeek publishes language-model weights and runs an inference API. Kuaishou’s Kling provides a separate example from video generation. For each release, the training artifacts, weights, license, and hosted access determine how other teams can use it.
Chronological Entries (4)
Kuaishou released Kling 4.0, with support for multiple keyframes in a video request. A creator could specify visual constraints at several moments across a shot, giving the generator more reference points than a single image.
DeepSeek released V4-Pro and V4-Flash in preview with both downloadable weights and hosted API access. Users could experiment through an endpoint or inspect and deploy the weights through their own serving stack.
Alibaba’s Qwen team described Qwen3-VL as a family with both dense and sparse expert variants. Text, images, and video can be interleaved in one context window, so the same model can answer a question that refers to both a document and a clip.
Direct Relations (0)
No explicit outgoing relations yet.