Tencent has completed a key part of its Hy4 release by publishing downloadable weights for the full flagship model. The company announced Hy4 on November 4, but its original communication said the weights and system-card material would follow. The official model repository now lists three distinct artifacts: instruction-tuned Hy4, an FP8 instruction checkpoint, and a pretrained Hy4 base. That is a material open-weight milestone, not a second launch of the already reported, smaller Hy4 flash. Primary source ↗
The headline numbers need a component-level explanation. Tencent describes Hy4 as a multimodal Mixture-of-Experts model with a 770-billion-parameter language backbone and 49 billion active language parameters per token. Its model table separately names a 1.23-billion-parameter vision transformer and a 3.3-billion-parameter DFly2 speculative-decoding drafter. Treating all those figures as an interchangeable model size, or interpreting 49B active as the total weight storage requirement, would mislead anyone planning hardware.
Hy4 supports text, images and video understanding and advertises a context ceiling of 1,048,576 tokens. Tencent uses sparse attention with index reuse and a shared-cache drafter to reduce some long-context serving costs. The official deployment instructions cover vLLM and SGLang, including multi-GPU FP8 recipes. Availability of those recipes does not mean a single consumer GPU can host the full checkpoint or serve a million-token session affordably.
The licensing distinction is equally important. Tencent labels the release with its own Tencent Hy License. Downloadable weights are real, but downstream commercial, redistribution and usage decisions must be made from the actual license terms and any applicable conditions. It would be incorrect to describe this release simply as Apache-2.0. The base checkpoint is also not an instruction-tuned assistant: prompt formatting and expected behavior differ.
Tencent reports strong results on computer use, office workflows and research tasks, including internal comparisons against other open models. The company itself identifies remaining gaps versus top proprietary systems. These results are useful leads for evaluation, not independent performance certificates. Open Model Weights has not reproduced Tencent's claims, and comparisons need exact checkpoint, harness, tool configuration, cost and inference precision.
The most important new opportunity is reproducibility. Researchers can now inspect the full artifact lineage, evaluate model and vision components, test deployment behavior and measure performance on their own workloads rather than infer everything from a hosted API. The next trustworthy milestone is an immutable repository snapshot and rights review, followed by measured end-to-end costs and independent task results. Until then, Hy4's open release and its performance assertions should be reported separately.
What this changes for the evidence layer
Newly downloadable Hy4 instruct, Hy4-FP8 and Hy4 base checkpoints are a substantive change from the November 4 API-led announcement and distinct from the previously reported smaller Hy4 flash release. Verify immutable repository revisions, complete model components, Tencent Hy License terms, actual weight tensors and runtimes before accepting new canonical registry records. No OMW benchmark run has been performed.
Tencent Hy — official Hy4 open-weight model card
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗