01 Trigger Event
Early on November 21, Tencent’s Chief AI Scientist Yao Shunyu and the Hunyuan team held a text-based AMA in an overseas social media community. Asked why ByteDance, Kimi, and other models appear to be advancing toward 3T or even 5T parameters while Tencent stopped at 700B, the team responded that Hy3 and Hy4 have demonstrated that smaller models can outperform larger ones at lower cost.
The same answer, however, acknowledged that Hy4 still has substantial room for improvement, particularly in post-training. The team is investing in large-scale reinforcement learning, user experience, and model safety, while planning to train larger models.
This is worth reading, but it is not a signal that Tencent has abandoned scaling. More precisely, Tencent has demoted scaling from the default path to one tool; it has not stopped expanding.
02 What It Really Means
The issue is not whether the scale should be 700B, 3T, or 5T. The real question is what scaling actually solves.
The Hunyuan team’s implicit judgment is that there is an optimal point between model capability and deployment cost. Beyond that point, adding parameters may yield limited benefits while directly increasing the complexity of training, inference, and serving. If Hy3 and Hy4 can indeed outperform larger models while using fewer parameters, the threat extends beyond competing products. It challenges the capital-allocation model of scaling parameters first and searching for use cases later.
“Cheaper,” however, is a relative claim; it does not establish a measurable cost-performance moat. The original response disclosed no parameter counts, benchmarks, latency, throughput, KV cache footprint, or API pricing. I have not run Hy3 or Hy4 internally, nor have I seen complete model cards, so the team’s claims cannot yet be treated as verified conclusions.
The potential moat is not a particular model name, but whether three conditions can hold simultaneously:
- Frontier capabilities do not materially collapse as parameter scale is reduced;
- RL and post-training continue to close capability gaps rather than relying only on static benchmarks;
- Deployment economics can support real-world traffic rather than merely laboratory figures.
If all three hold, 700B can become a better operating point. If post-training fails to take over, “smaller models win” will remain only a temporary benchmark advantage.
03 Historical Analogy / Structural Comparison
This resembles AWS’s strategy around 2014 more closely than the launch of the iPhone in 2007.
The iPhone introduced a new computing paradigm. AWS demonstrated that infrastructure could replace enterprises’ direct purchases of expensive hardware through higher utilization, lower unit costs, and lighter-weight delivery. Hunyuan’s current emphasis on smaller, cheaper models is similarly competing for an advantage in unit economics.
There is, however, a critical difference. AWS did more than claim that cloud was cheaper than servers; it built a complete product structure around instance types, APIs, regional availability, and customer migration paths. A model provider seeking to build a moat through efficiency must similarly penetrate the entire stack:
- Can routing automatically select models according to task difficulty?
- Can KV cache and prefix caching reduce the cost of long-context workloads?
- Can speculative decoding, quantization, MoE activation, and batch serving improve effective throughput?
- Can model updates maintain quality without allowing regression testing and eval costs to consume the gains?
Parameter count is therefore not the strategy itself. The real strategy is controlling the total cost of every successful task, including tokens, retries, latency, human acceptance, and failure costs. This is why model routing and token gateways are valuable: they translate differences in model capability into customer-procurable SLAs and bills.
Tencent also retains a trump card: continuing to train larger models. This indicates that the company does not believe efficiency alone will impose a hard ceiling on capability. It is closer to buying two options—using smaller models to protect the cost curve while using large models to preserve the capability ceiling.
04 What It Means for AI Builders
The most important task this week is not to move all production traffic to Hunyuan. It is to recalculate your model portfolio.
First, map “each task,” not merely “each API call.” Track model unit prices, input and output tokens, cache hits, tool-call counts, retry rates, end-to-end latency, and human rework. For agent workflows, the final item is often more important than the token bill.
Second, establish capability tiers rather than betting on a single model. Route frequent, clearly defined tasks to inexpensive small models, then escalate complex reasoning to a frontier model. The potential arbitrage window lies in using routing behind a single API to absorb differences in capability. I have not internally benchmarked Hy3 or Hy4, so no unverified escalation thresholds can be offered.
Third, post-training should be part of the product roadmap, not merely the vendor roadmap. If your differentiation comes from workflows, tool use, domain data, or feedback loops, retain control of RL data, reward design, eval harnesses, and regression suites. Hunyuan’s emphasis on post-training highlights how the remaining value between foundation models and vertical applications is migrating toward the post-training layer.
Fourth, do not treat “plans to train larger models” as guaranteed supply, and do not ignore those plans merely because the current scale is smaller. Model supply may still produce capability step changes. A router must support hot-plugging new models, fallback mechanisms, and automatic redistribution after price changes. Otherwise, efficiency is only a static architectural advantage.
For independent AI product entrepreneurs, Tencent’s potential advantage lies not in a single benchmark but in its combination of models, cloud infrastructure, distribution, and enterprise relationships. The opportunity may instead sit in the cross-model layer: whoever can provide verifiable cost, quality, and latency optimization across different models may become the AWS control plane of the AI era rather than an appendage to any single model.
05 Counterargument / Risks
I may be underestimating the disruptive power of larger models.
The most cautionary element of this AMA is not the team’s self-assessment that Hy4 can still improve. It is the combination, within the same conversation, of an emphasis on efficiency and an admission that scale will continue to expand. If scaling law has not genuinely failed, today’s small-model advantage may be temporary. The next change in training, post-training, or test-time compute could reopen the gap.
A second risk is the disconnect between benchmarks and real user experience. The team says Hy3 and Hy4 have demonstrated that smaller models can outperform larger ones, but disclosed no task distributions, sampling parameters, agent scaffolds, inference budgets, or cost definitions. This is especially important for agent products, where a model’s raw capability and a system’s completion rate may not be the same. I have not run this system internally and therefore cannot confirm its claimed advantages.
Finally, Tencent owns WeChat, Tencent Cloud, and enterprise distribution. Even if its model capability reaches only the industry’s leading tier, it may reduce customer-acquisition costs through default placement, bundling, and channel subsidies. This would undermine the assumption that efficiency-driven price advantages are a durable moat: price advantages may not survive independently because competitors ultimately sell distribution and ecosystems rather than a quoted price per million tokens.
My conclusion is not that Tencent has found the correct post-scaling path. It is that the company is temporarily refusing to treat parameter count as the only strategy. The most dangerous mistake, however, would be mistaking restraint for victory. The decisive evidence will be Hy4’s complete technical report, reproducible benchmarks, and API economics supported by real-world traffic.