Tencent Hunyuan has released and open-sourced Hy4 Preview. This new generation flagship MoE large model has a total parameter of 770B, 49B per token activation, and the native context length has been increased to 1MB. Rather than simply chasing rankings, Tencent places greater emphasis on real productivity tasks such as coding, office work, game development, and scientific research.
770B MoE presses 1M of context
The Hy4 preview backbone contains 78 layers, of which 77 use MoE, each layer is configured with 256 routing experts and 1 shared expert, and each token activates the Top-8 routing expert. Model reduction adds a layer of MTP for speculative decoding.
The attention architecture uses Gated DSA, combined with IndexCache multiplexing cross-layer sparse indexes, and introduces iHC in residual paths. Tencent states that compared to the previous generation, Hy4 preview simultaneously expands model scale, context, and training data.
Shift from benchmarking to real productivity
Tencent has listed AI programming, office analytics, game development, and scientific research as core scenarios for Hy4 Preview, with training data incorporating real-world tasks from internal software engineers, financial analysts, and security experts, and continues to collaborate with CodeBuddy and WorkBuddy.
In Tencent's internal blind test, 163 experts evaluated 203 engineering tasks. The Hy4 preview averaged 2.99/4, with a win rate of 46.8% against GLM 5.3 and 51.2% against Kimi K3. This set of data further shows that Tencent is pushing model competition toward complex task completion, not just standardized benchmarks.
Open-source weight seizes the Agent ecosystem
Hunyuan Hy4 open-source synchronously covers the original weights and FP8 versions, uses the Apache-2.0 license, supports deployment via vLLM and SGLang, and calls via OpenAI-compatible APIs. The model has been launched on platforms such as Hugging Face, ModelScope, and GitCode.
The total parameter scale of the 770B also means that the Hy4 preview is more suitable for server-level deployment rather than ordinary personal devices. For Tencent, open source is truly competing for the model foundation positions for AI Agents, enterprise office, and development tools, while CodeBuddy and WorkBuddy provide the entry point for continuously returning real task data.
The Preview identity also leaves clear room for improvement. Tencent confirms that the current version still faces issues such as overly long thinking for complex tasks and excessive self-validation. Whether the official version of Hy4 can maintain long-context and complex task capabilities while reducing inference costs will determine whether it can move from an "open-source frontier model" to large-scale production environments.
