Tencent releases Hy4 flash weights in three variants under a custom license with revenue and user caps
Tencent has released downloadable weights for Hy4 flash in three variants: instruction-tuned, FP8 and base. The checkpoints ship under Tencent Hy License 1.0, which requires a separate license above US$50 million in revenue or 100 million monthly active users. Tencent says the larger Hy4 model's weights are still being prepared.
Tencent has released downloadable weights for Hy4 flash in three variants, following the model's November 16 API announcement. IT Home reported the company's weight-release announcement on November 18. The checkpoints are published under Tencent Hy License 1.0, which is not Apache 2.0 and sets commercial thresholds.
Three variants
Tencent's official model card lists:
- Hy4 flash: the instruction-tuned model.
- Hy4 flash-FP8: the quantized instruction-tuned counterpart.
- Hy4 flash-base: a pretrained base model, intended for further training rather than as a drop-in replacement for the chat-ready variants.
Pandaily checked all three official Hugging Face repositories. The weight-index files referenced available shard listings, and anonymous byte-range requests returned weight-file data for each variant. These checks confirm download availability. They were not full-checkpoint downloads or successful inference tests.
Checkpoint size
The standard checkpoint's index records approximately 440.5 GB of tensor data. The FP8 checkpoint records approximately 287.1 GB. These are weight-data sizes, not minimum GPU-memory requirements. Quantization reduces storage, but this remains a large model, not a consumer-device release.
Deployment requirements
Tencent provides dedicated vLLM and SGLang container images, with examples that use eight-way tensor parallelism. Its Transformers instructions point to a model-support pull request, so an ordinary released package does not yet appear to support the architecture.
The vLLM recipe calls for host RAM because Engram tables are offloaded to CPU memory by default. Its starting configuration uses an 8,192-token context. The model's advertised 1M-token capacity is not a guarantee that every hardware configuration can serve that length. Published memory hints are estimates, not measured minimum requirements.
License terms
Tencent Hy License 1.0 states explicitly that it is not Apache 2.0, despite being derived from it. Section 10 requires a separate written license in two cases:
- Model-as-a-service operators exceeding US$50 million in aggregate worldwide revenue over a consecutive 12-month period.
- Use of the model in a product or service with more than 100 million monthly active users.
A narrowly defined internal-use exception applies when the model and its outputs are not provided to third parties.
Tencent says the weight release for the larger Hy4 model is still being prepared.
What this means
The release gives teams deployable checkpoints for self-hosting and, through the base variant, a starting point for further training. The tooling requirements set the practical bar: eight-way tensor parallelism in the documented examples, CPU memory for Engram tables, and a Transformers path that still depends on an open pull request.
The license decides who can use the weights freely. Tencent's own wording rules out calling this Apache 2.0, and the revenue and monthly-active-user thresholds mean large service operators need a written agreement. The release is best described as open-weight under Tencent's stated conditions, not unrestricted permissive open source.
Sources: IT Home (weight-release announcement), Tencent's official project page, model card and Hy License, and the vLLM deployment recipe.
Related Articles
Tencent releases Hy4 flash: 220B MoE, 12B active parameters, 1M-token context
Tencent has released Hy4 flash, a multimodal Mixture-of-Experts model with 220B total parameters, 12B active per token, and a 1,048,576-token context window. Tencent calls it its most cost-effective Hy model to date and claims a per-task cost of 1/20 that of Hy4 on its internal WorkIndex benchmark.
Tencent releases Hy4 Flash Base: 220B-parameter MoE, 12B active, 1M-token context
Tencent has published Hy4 Flash Base on Hugging Face: a 220B-parameter mixture-of-experts model with 12B active parameters per token, text and image input, and a 1M-token context. Tencent calls it its most cost-effective Hy model to date.
GitHub Copilot team customizations now generally available for organization teams
GitHub announced that team customizations for GitHub Copilot are generally available for organization teams. The feature lets teams share the Copilot plugins they rely on without asking every member to repeat the same setup.
Google adds "temporary threads" to AI Mode in Search, detaching chats from your account
Google is adding "temporary threads" to AI Mode in Search, a chat option that detaches the session from the user's account. The conversations won't resurface in search history or recommendations. The feature is rolling out globally on desktop and mobile across all supported languages.
Comments
Loading...