Hy4

Tencent's Hy4 has moved from preview to official general availability, and the GA release is the version that adds visual understanding for the first time. Tencent's release description claims substantial improvements over Hy4 Preview, with particularly strong gains in agentic capability and long-horizon productivity tasks, so teams should evaluate the GA variant on their own workloads rather than assuming preview-level behavior carries over. The GA model and the earlier preview remain distinct catalog entries, and the upgrade case is capability rather than configuration: the GA entry accepts both text and image input where the preview entry was text-only, while output stays text. The vendor's agentic and visual claims are self-reported without supplied benchmark tables, so independent evaluation is still needed before moving production traffic.

Tencent released and open-sourced Hunyuan Hy4 preview on 28 August 2026 as a 770B-parameter MoE model from the Tencent Hy Team that activates 49B parameters per token. Weights ship in BF16 as tencent/Hy4-preview and in FP8 as tencent/Hy4-preview-FP8 under the Apache 2.0 license, published on Hugging Face, ModelScope, GitCode and CNB, with vLLM and SGLang serving recipes and prebuilt container images. In a blind side-by-side evaluation, 163 internal experts rated outputs on 203 engineering tasks, with Hy4 preview averaging 2.99 against GLM 5.3 at 2.92 and Kimi K3 at 2.94. Tencent also documents known limitations, noting the model can reason longer than necessary on complex tasks and tends to over-verify its own work, framing the release as a deliberate early ship.

IntuitionLabs analyzed Tencent Hy4 preview, released by the Tencent Hy Team on August 28, 2026 as an Apache 2.0 open-weight MoE model with 770 billion total parameters and 49 billion activated per token. Its weights are published as both a full-precision Hy4-preview checkpoint and an Hy4-preview-FP8 quantized checkpoint, distributed through Hugging Face, ModelScope, GitCode and CNB. The analysis notes that Hy4 preview supports roughly a one-million-token context window and includes a built-in multi-token prediction layer for speculative decoding. It also cautions that Tencent's blind comparative evaluation was run internally and self-selected, so the model's benchmark standing reflects a snapshot of an internal test suite rather than a third-party-audited ranking.

Tencent's Hy Team published Hy4 preview on Hugging Face as a new-generation Mixture-of-Experts flagship model with 770B total parameters and 49B activated per token. The backbone has 78 layers, where the first uses a dense FFN and the remaining 77 use MoE with 256 routed experts plus one shared expert, activating the top-8 routed experts alongside the shared expert. A native MTP layer of 10B total parameters with 0.7B activated is built in for speculative decoding, and attention uses Gated DeepSeek Sparse Attention with IndexCache for cross-layer sparse index reuse, inspired by DeepSeek and GLM. The residual pathway uses iHC identity Hyper-Connections, with 1M context length, 120,832 vocabulary size, and Apache 2.0 licensing.