Tencent: hy4

Tencent: Hy4 is a flagship open-weight model from Tencent Hunyuan, released on August 28, 2026 under the Apache 2.0 license. It is a Mixture-of-Experts architecture with 770B total parameters and 49B activated parameters per token, built from 78 Transformer layers — the first a standard feed-forward layer and the remaining 77 MoE layers, each with 256 routed experts and 1 shared expert, activating top-8 routed experts plus the shared expert per token. Hidden size is 6144, feed-forward intermediate size 18432, and MoE intermediate size 2048, with a vocabulary of 120832 tokens. It accepts text input and returns text output, supports up to 1M tokens of context with a 64K-token output limit. The architecture adds Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, identity hyper-connections that expand residual paths to 4 residual streams, and a built-in native multi-token prediction layer (10B total, 0.7B activated parameters) for speculative decoding. Two response modes are offered: a high-effort reasoning mode and a no_think direct response mode. Deployment recipes for vLLM and SGLang are provided with 8-way tensor parallelism and an FP8 variant. Hy4 is positioned for real productivity work and was tuned on data drawn from Tencent internal team workflows. It suits software engineering, office and document generation, game prototype development, and scientific research. Tencent describes it as an early, imperfect version with room for improvement in pre-training and post-training, noting known issues such as overly long reasoning chains and excessive self-verification. The API model ID is hy4-preview, accessible via Tencent Cloud International TokenHub and OpenRouter, and the model is integrated into CodeBuddy, WorkBuddy and Yuanbao. Weights are available on Hugging Face, ModelScope, GitCode and CNB.