Tencent HY 2.0

What it is

Tencent HY 2.0 is the flagship model Tencent developed in-house, released on Friday 5 December 2025 ^[1]. It comes in two versions, HY 2.0 Think and HY 2.0 Instruct ^[1]. The split reflects a division of labour that has become common: one variant is trained to work through a problem at length before answering, while the other answers directly, which is cheaper and faster for ordinary product use. Tencent put the model into its own AI-native products rather than offering it only as an API ^[1].

Architecture

HY 2.0 uses a mixture-of-experts (MoE) architecture with 406 billion total parameters, of which 32 billion are active ^[1]. In an MoE model the network is divided into many separate expert subnetworks, and a routing step sends each token to only a few of them. The model can therefore hold far more parameters than it uses on any single token, so the compute cost of a forward pass stays closer to that of a much smaller dense model. The context window is 256K tokens ^[1] — the amount of text the model can keep in view at once, enough to hold a long technical document together with its code.

Training and reasoning

Tencent says the Think version improves on the previous generation in both its pre-training data and its reinforcement-learning strategies, and reports strong results on complex reasoning in mathematics, science and coding ^[1]. The company credits a dual-stage reinforcement-learning approach that combines reinforcement learning with verifiable rewards (RLVR) and reinforcement learning from human feedback (RLHF) ^[2]. The two do different jobs. RLVR trains against checks that can be run automatically — a maths answer compared with a reference answer, code that either passes its tests or does not — so the feedback is cheap and unambiguous wherever such a check exists. RLHF shapes responses using human judgements of quality, which covers what no automatic checker can score.

Reported benchmark results:

• HY 2.0 Think scored 73.4 on IMO-AnswerBench, a mathematics benchmark ^[2].
• On SWE-bench Verified, which tests whether a model can resolve real issues in software repositories, the score rose from 6.0 for the previous model to 53.0 ^[2].
• On Tau2-Bench, which evaluates a model across multi-turn tasks that involve calling tools, the score rose from 17.1 to 72.4 ^[2].

Both of the last two figures start from a low baseline, so the absolute gains are large: the earlier generation was close to failing such tasks, while HY 2.0 Think solves a substantial share of them.

Where it can be used

HY 2.0 was integrated into Tencent's AI-native products Yuanbao and ima, and made available through Tencent Cloud APIs ^[1]. Dataconomy reported that Tencent planned to open-source related technologies and models ^[2].

Place in the line

The successor to HY 2.0, Hy3, arrived in July 2026 with 295 billion total parameters, fewer than HY 2.0's 406 billion ^[3]. The next step in the line was therefore smaller in total parameters rather than larger.