Hy3 Convergence: Tencent AI Shifts from Internal Rivalry to Single Strategy

Photo by yating yang / Pexels

---

The most valuable information in this article is that Tencent, through the convergence of the Hy3 model foundation, integrated its previously scattered AI product lines into a unified strategy. Essentially, this is a "portfolio optimization" move—cutting redundant assets with low Sharpe ratios and high correlation, concentrating resources on a single model with high expected returns and low friction costs. This aligns better with the "minimize slippage + maximize Sharpe" principle in high-frequency trading than parallel multi-model approaches.

Comparison: Parallel Multi-Models vs. Single Model Foundation

Tencent's previous AI product layout can be likened to a quantitative trading team running 10 different strategies simultaneously, each using different factor libraries, risk models, and even data sources. Yuanbao, ima, WorkBuddy, CodeBuddy, QClaw, Marvis, Xiaowei, QBot... these products came from different teams, had different underlying models, and targeted different scenarios. This "diversification" seemed to offer the benefit of "risk diversification through multiple strategies," but in reality, it brought three fatal problems:

1. High Correlation: The training data for different models heavily overlapped (all text, code, and dialogues within the Tencent ecosystem), but architectural differences led to inconsistent outputs. This "pseudo-diversification" often leads to simultaneous drawdowns in live trading—for example, multiple models hallucinating against enterprise clients at the same time, causing brand damage.

2. High Slippage Costs: Each product independently maintained inference services, API interfaces, and model update processes, equivalent to redundantly deploying infrastructure across multiple trading accounts. Estimating by compute costs, adding each independent model increased inference latency by 15%-25%, while operational complexity grew linearly. "Ideal returns" in backtests were largely eaten up by live trading slippage.

3. Low Sharpe Ratio: Sharpe Ratio = (Expected Return - Risk-Free Rate) / Volatility. The "expected return" of parallel multi-models is the average effect across dispersed scenarios, but "volatility" rises significantly due to model inconsistency and varying answer quality. A typical example: Users getting different answers from Yuanbao and QBot for the same question, leading to decreased trust and increased volatility in product usage rates. The overall Sharpe ratio might be below 0.5.

The convergence after the Hy3 release is essentially merging multiple strategies into one "core strategy." This strategy uses the unified Hy3 model foundation but allows different products to adapt to specific scenarios via fine-tuning (LoRA) or prompt engineering. This is similar to the quantitative trading practice of using a "Core Factor Model + Conditional Weights": the same factor library applies to all assets, but factor exposures are adjusted for different asset classes.

Quantitative Advantages of Convergence: From Sharpe Ratio to Live Performance

Let's speak with data. Assume Tencent internally had 5 independent models (A, B, C, D, E), each with a Sharpe ratio of 0.3, but with a mutual correlation of 0.8. If you make an equal-weight portfolio, the Portfolio Sharpe Ratio = 0.3 / sqrt(0.8) ≈ 0.335, a limited improvement, and portfolio volatility is only slightly lower. But if all model resources are concentrated on Hy3 to train a larger, more unified model, assuming its Sharpe ratio reaches 0.6 (due to more data, fuller training, and better inference consistency), the overall Sharpe ratio doubles directly.

More importantly, "slippage" decreases in live performance. Here, slippage refers to latency, cost, and inconsistency during model inference. After unifying the Hy3 foundation, inference services can share a single compute pool, making load balancing more efficient. According to Tencent's published Hunyuan Large Model technical report, Hy3 adopts an MoE architecture, improving inference efficiency by over 40%. This means the cost (time + compute) for each product calling the model drops, equivalent to "lower commissions" in trading, directly boosting net returns.

Additionally, convergence reduces "Value at Risk" (VaR). With parallel multi-models, the worst-case scenario is a severe hallucination or crash in one model affecting all dependent products. With a unified foundation, one model version controls output quality for all products, allowing quick rollback to the previous version if issues arise. This "centralized risk management" is more controllable than decentralized "fighting separately."

[!info] Key Judgment

Convergence isn't "abandoning diversity," but "removing redundant diversity." Retain scenario fine-tuning capabilities, but eliminate inconsistencies at the model layer. It's like retaining multi-asset allocation in trading but using the same risk model to manage all assets.

Tencent's Path to Convergence: The Cost of Going from "Nine Sons" to "One Son"

Convergence isn't without cost. The R&D resources, team building, and brand awareness Tencent invested in multiple models over the past few years are sunk costs. Abandoning certain models means abandoning specific product positions. For example, CodeBuddy is a code assistant for developers, WorkBuddy for office scenarios, and QClaw for government affairs. If all switch to Hy3 fine-tuned versions, there might be initial performance declines—because Hy3's general capabilities might not be as precise as a model trained specifically for code.

But there's a classic conclusion in quantitative trading: In low signal-to-noise environments, simple models + sufficient data usually outperform complex models + small data. The data volume for coding scenarios is far smaller than general text; training a dedicated code model might lead to overfitting due to data sparsity, whereas Hy3, as a general large model, improves generalization in coding tasks through superior scale of pre-training data, potentially surpassing specialized models. Tencent's choice of the Hy3 foundation is essentially betting that "scale effects" outweigh "customization effects."

Original link: https://www.tmtpost.com/8063089.html