Tencent offers instruct, FP8 instruct and pretrained base weights for Hy4 Flash, with text/image input, text output and a stated context up to one million tokens. Its five-layer DFly2 drafter reuses the target model’s KV cache and up to 2,048 sparse-indexer positions instead of maintaining another cache. Tencent says it does not participate in prefill.

Tencent’s internal WorkIndex reports 42.1 against Hy4 Preview’s 39.7 at one-thirteenth the average task cost. Those figures belong to Tencent’s setup. Its Hy License is not Apache 2: qualifying model-as-a-service businesses crossing $50 million worldwide annual gross revenue or 100 million monthly active users face a separate written-license requirement. The internal-only exception requires models, derivatives and outputs to remain unavailable to third parties. Review the actual license for the intended distribution model.

huggingface.co