Is Tencent’s Large Model Lagging? How Hy3‑preview Propels It Into the Top Tier

While Chinese large‑model development is accelerating, Tencent has been perceived as moving more slowly due to fragmented AI research groups, mismatched engineering and model capabilities, and high coordination costs. In late 2023, chief AI scientist Yao Shunyu joined Tencent and advocated breaking down departmental walls, leading to the dissolution of a decade‑old AI Lab.

Complete Rebuild

About three months ago the Hunyuan team tore down and rebuilt the entire stack: data processing for pre‑training, reinforcement‑learning reward mechanisms, and the underlying compute infrastructure. The effort produced Hy3‑preview, a model with 295 B total parameters that activates 21 B parameters per request and adopts a fast‑slow hybrid expert architecture.

The fast‑slow mechanism works as follows: for simple daily conversations the model quickly invokes a small set of parameters to generate an answer; for complex logical reasoning it slows down, engaging deeper network layers and more parameters. Coupled with a 256 K context window, this makes Hy3‑preview the most capable Hunyuan model to date.

Benchmark Excellence

Hy3‑preview excels on scientific and mathematical benchmarks such as FrontierScience‑Olympiad and IMOAnswerBench, achieving the highest domestic scores in Tsinghua University’s 2026 spring math PhD qualification exam and strong results in the 2025 CHSBO biology competition. The team also introduced two new benchmarks, CL‑bench and CL‑bench‑Life, to evaluate the model’s context‑learning ability, which show significant gains after the new evaluation regime.

Agent‑Level Capabilities

The most noticeable improvement is in the Agent ability. The model can write code, retrieve information, and invoke external tools to complete tasks, effectively turning high‑level instructions into concrete outcomes. For example, given a prompt to create a small‑planet resource‑gathering game, Hy3‑preview outputs a complete design with gameplay mechanics, graphics, and sound suggestions. It also summarizes data reports into polished presentations, generates full WeChat mini‑program code from a single requirement, and extracts actionable items from chaotic group chats.

Real‑World Development Performance

In comprehensive evaluations such as SWE‑Bench Verified, Terminal‑Bench 2.0, BrowseComp, WideSearch, ClawEval, and WildClawBench, Hy3‑preview shows a large margin over its predecessor and competitive results against leading open‑source models. Internal test suites covering backend engineering (Hy‑Backend), user‑interaction (Hy‑Vibe Bench), and high‑difficulty software engineering tasks (Hy‑SWE Max) also confirm strong performance under pressure.

Open‑Source Release and Efficiency Gains

The model weights and code are fully open‑sourced on GitHub, HuggingFace, ModelScope, and GitCode, with support for vLLM and SGLang inference frameworks. Continuous optimizations in inference kernels, quantization algorithms, and framework integration yield a 40 % increase in inference speed and substantially lower operating costs compared with the previous generation.

Overall, the ground‑up reconstruction of the Hunyuan model has elevated its underlying capabilities, allowing Tencent’s large language model to join the first‑tier competitors in the global AI landscape.