Tencent Releases Hy4 Preview Open Model Weights

Tencent's Hugging Face documentation describes a 78-layer backbone. The first layer uses a dense feed-forward network; each of the remaining 77 layers contains 256 routed experts and one shared expert. For each token, the router selects the top eight routed experts alongside the shared expert.

The model card lists Gated DeepSeek Sparse Attention, or Gated DSA, with cross-layer sparse-index reuse, plus four residual streams. Tencent also includes a native multi-token-prediction, or MTP, layer with 10 billion total parameters and 0.7 billion activated parameters, intended for speculative decoding.

The distinction between total and activated parameters matters operationally, but it does not eliminate deployment costs. The vLLM Hy4 serving recipe lists a BF16 weight footprint of 1,848 GB and an MXFP8 footprint of 924 GB. Its reference configurations identify at least 16 NVIDIA B200 GPUs or eight B300 GPUs for BF16 weights plus KV cache. Long-context workloads add substantial KV-cache pressure beyond model-weight memory, a general constraint for models with million-token advertised windows.

According to Tencent Cloud International's August 28 FAQ, the API model ID is hy4-preview, with a 960,000-token maximum input and 64,000-token maximum output within the stated 1 million-token context budget. The same FAQ lists structured output, function calling, caching, and preserved-thinking capabilities.

TechNode reports API pricing of $0.834 per million input tokens and $2.501 per million output tokens, and reports that WorkBuddy and CodeBuddy are offering two weeks of free access. Tencent Cloud International characterizes the model as an early preview rather than a generally available production release, noting that pre-training and post-training remain subject to improvement.

For engineering teams, that status makes measured evaluation more useful than parameter-count comparisons. Teams evaluating comparable preview models commonly need workload-specific tests for tool-call validity, structured-output conformance, retrieval performance at long context, latency, and cost under realistic concurrency.

Tencent's model card describes Hy4 Preview as targeting software engineering, game development, finance analysis, and security-related productivity work. TechNode reports that, in an internal blind evaluation involving 163 experts and 203 engineering tasks, Hy4 Preview scored 2.99 out of four, versus 2.92 for GLM 5.3 and 2.94 for Kimi K3.

TechNode also reports Tencent's claim of a 31.8% end-to-end throughput increase versus a baseline after using the model to optimize parts of its training and inference systems. Those figures are internal results rather than independently reproduced benchmarks. External testing, especially across long-context and agentic coding workloads, remains necessary to establish how the model compares in production-like environments.