No. 124
2026-11-20
the AI pulse — brewed, not raw
 AI CRITIQUE
Mavi gökyüzünde dağınık bulutların önünde, mavi cam cepheye sahip ve üzerinde "Tencent" logosu belirgin şekilde yer alan yüksek, modern bir bina.

Tencent releases Hy4 flash model weights for local deployment

Read the article    English (automatic translation)

Summary

Tencent has released the weights of its Hy4 flash multimodal model for local deployment, offering standard, FP8, and base versions. The model features approximately 224 billion total and 12 billion active parameters, with a 1 million token context window. This release came two days after the API launch.

Why it matters

Making a large MoE model accessible for on-premise use is a significant step for enterprises that cannot send data to cloud APIs. The 1 million token window and text-image modality place the model in the same league as leading Western releases. However, the Tencent Hy License is proprietary. This indicates that while the model is 'open-weight,' it is not truly open source. Without Tencent's permission, the community's ability to redistribute or build upon the model is limited.

Highlights

  • 224 billion total / ~12 billion active parameters, 1 million token context, text + image input
  • Three downloadable versions: standard, FP8, and base model for fine-tuning
  • Released under the proprietary Tencent Hy License instead of a permissive open-source license

Critical take

The 'local deployment' headline is somewhat misleading. Even with only 12 billion active parameters, a 224-billion-parameter MoE model requires multi-GPU server hardware. Therefore, 'local' here refers to an enterprise data center, not a developer's laptop. Coupled with the proprietary license, the release's actual openness is far more limited than the framing suggests.

This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.
Back to the issue