最新模型
Tencent Releases Official Hy4 Weights; vLLM Deployment Requires a New Image or Nightly Build
On November 20, Tencent released the instruction-tuned, FP8, and pretrained base weights for Hy4, moving its flagship multimodal model into the self-hosting phase. The official deployment guide provides a speculative decoding configuration, but the million-token context window, hardware requirements, and custom license each require separate review.
Tencent released the official Hy4 weights on November 20, including the instruction-tuned model, an FP8 quantized version, and a pretrained base model. The release turns the capabilities announced earlier into an implementation that can be downloaded and independently validated. The base model has not undergone instruction tuning, so expectations for using a chat model do not apply directly. Official model card
Hy4’s language backbone uses a mixture-of-experts architecture, with 770 billion total parameters and about 49 billion activated per token. Of its 78 layers, 77 use expert feed-forward networks. Each layer selects eight of 256 routed experts, alongside one shared expert. Its attention combines Gated DSA with cross-layer sparse index reuse, and it processes images and video frames through a vision encoder. These design choices mean deployment assessments must account for weight storage, expert computation, and long-context cache together; activated parameter count alone cannot be used to estimate GPU memory. Official repository
Inference configuration is a practical focus of this release. The vLLM documentation states that version 0.34.0 does not yet register the Hy4 architecture, so deployment requires a dedicated image or a nightly build from November 20 onward. The example uses tensor parallelism across eight GPUs, together with the hy_v4 tool and an inference parser. The built-in DFly2 draft model reuses the main model’s KV cache and sparse indices, avoiding a separate draft cache. However, the recipe starts with a 131,072-token window and explicitly says that full million-token requests have not yet been validated; the model’s stated maximum context length should not be treated as a guarantee of serving capacity. vLLM deployment recipe
The license also requires a separate review. The repository uses Tencent Hy License 1.0, which the documentation says is adapted from Apache 2.0 and adds commercial terms for large-model serving businesses. Engineering teams should pin the image version, measure time to first token, speculative decoding acceptance rate, and peak memory usage for long-context workloads, then assess deployment costs. The official capability and performance claims still need independent reproduction. License text