Tencent has released and open-sourced Hy4 preview, a new flagship large language model from its Hunyuan team that the company is positioning at the open-source frontier. The weights landed on Hugging Face, ModelScope, GitCode, and CNB on August 28 under the permissive Apache 2.0 license, alongside an FP8-quantized variant for cheaper inference.
The launch intensifies the competition among Chinese labs shipping open-weight frontier models, a field that currently includes Z.AI's GLM series and Moonshot AI's Kimi models. Bloomberg reported that Tencent claims the new model outperforms both rivals in internal testing — a claim that, like all vendor-run evaluations, deserves scrutiny. For more context on this story, see our ongoing latest AI developments.
The Headline Numbers
Hy4 preview is a Mixture-of-Experts (MoE) model with 770 billion total parameters, of which 49 billion are activated per token. According to the official model card published on the Tencent-Hunyuan GitHub repository, the backbone has 78 layers, 256 routed experts plus one shared expert, and activates the top-8 routed experts for every token. A native multi-token prediction (MTP) layer — 10 billion total parameters, 0.7 billion activated — is built in for speculative decoding.
The context window is the other marquee specification: 1 million tokens, a length that lets the model process entire codebases, long legal documents, or hour-long meeting transcripts in a single pass.
Architecture Borrowed From Rivals — and Built On Top of It
In an unusual degree of candor for a flagship launch, Tencent's documentation states that the attention module was "inspired by DeepSeek and GLM." Hy4 preview uses Gated DeepSeek Sparse Attention (Gated DSA) combined with IndexCache for cross-layer sparse index reuse, while the residual pathway employs identity Hyper-Connections (iHC) to expand inter-layer information flow.
The openness matters for developers evaluating the model: sparse attention is what makes a 1M-token context economically servable, since naive full attention at that length becomes prohibitively expensive. Both DeepSeek and Z.AI have shipped variants of this design in their own flagship models.
Built for Productivity, Tuned With In-House Experts
Tencent says it partnered with experts inside the company — software engineers, game developers, finance analysts, and security experts — and built training data around the work they actually ship. The model card highlights four focus areas:
- Software engineering: improved planning, debugging, and verification on long-horizon development tasks, plus gains in front-end "visual taste."
- Office and analysis: converting messy multi-file context into shareable documents, spreadsheets, and presentations, with stronger handling of equations and financial models.
- Game development: generating playable prototypes from a single prompt and working fluently with game engines across multiple refinement turns.
- Scientific research: claimed progress on AI research, molecular dynamics, condensed matter physics, and pure mathematics problems.
Tencent also says Hy4 preview was co-designed with its own products, CodeBuddy and WorkBuddy, so that model gains translate into product improvements.
What Tencent's Own Benchmarks Show
The most concrete comparison in the launch materials is a blind side-by-side evaluation Tencent ran with 163 internal experts rating outputs on 203 engineering tasks. In those tests, Hy4 preview scored an average of 2.99, slightly ahead of GLM 5.3's 2.92 (46.8 percent wins, 12.8 percent ties, 40.4 percent losses) and Kimi K3's 2.94 (51.2 percent wins, 7.9 percent ties, 40.9 percent losses).
Two caveats are worth stating plainly. First, these are Tencent's internal evaluations, not independent benchmarks; the company has not yet published third-verified results on public leaderboards. Second, the margins are narrow — a 0.05 to 0.07 point gap on a subjective expert-rating scale is within the range that different rating pools could easily flip.
Independent verification will come from public aggregations as the community runs the model through standardized coding, reasoning, and agentic suites in the coming weeks.
Shipping Early, With Known Bugs
Tencent is explicit that this is a preview. The model card lists known limitations, including spending longer than necessary reasoning through complex tasks and a tendency to over-verify its own work. The team says it would "rather ship early and hear what breaks," citing the approach that improved its Hy3 generation.
For deployment, the weights are available in both BF16 and FP8, with day-one support in vLLM and SGLang. Tencent publishes prebuilt Docker images for both, with recipes recommending tensor parallelism across 8 GPUs and MTP-based speculative decoding for latency-sensitive serving. A finetuning pipeline and the AngelSlim quantization toolkit ship alongside the model.
Why It Matters
The open-weight race in China has become a two-front war between Z.AI's GLM series and Moonshot's Kimi models, with DeepSeek, Qwen, and now a much larger Hunyuan entry crowding the field. Hy4 preview's combination of Apache 2.0 licensing, a 770B-parameter scale, and a 1M-token context raises the ceiling of what anyone can download and self-host — and the day-one vLLM and SGLang support lowers the barrier to actually running it.
For enterprises weighing open models against Western closed APIs, the practical question is no longer whether an open model can match closed performance on paper, but whether the surrounding tooling — quantization, serving, finetuning — is mature enough. Tencent is betting that shipping all of it at once, under a permissive license, is how Hunyuan gets back into the conversation.
Whether the internal benchmark claims hold up under independent testing will determine if Hy4 preview is remembered as a genuine frontier open-source release or another vendor benchmark in a crowded field. The weights are public now, so the community's verdict should not take long.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
