Tencent’s latest flagship AI model is not its largest. That may be the most interesting thing about it.
The company has introduced Hy3 preview, an open-source large language model with 295 billion total parameters and 21 billion activated parameters. That still places it firmly in large-model territory, but it also makes Hy3 preview more restrained than some of the industry’s biggest recent releases — and, notably, more compact than Tencent’s earlier flagship generation. Rather than treating scale itself as the headline, Tencent is presenting Hy3 preview as a model built around deployability, efficiency and real-world use.
That reflects a broader shift in the market. The first phase of the model race was largely defined by size: more parameters, larger training runs, more expensive infrastructure. But the commercial reality of the current market is different. Capability still matters, but so do latency, serving cost and whether a model can be deployed widely enough for pricing to matter. Hy3 preview appears to be Tencent’s answer to that shift.
Tencent says the model improves inference efficiency by 40%, while supporting long-context workloads up to 256K and targeting areas such as reasoning, instruction following, coding and agentic tasks. It is also being priced accordingly. On Tencent Cloud’s TokenHub, API access starts at RMB 1.2 per million input tokens and RMB 4 per million output tokens, with the company also highlighting lower deployment barriers for enterprise users.
The company’s latest briefing makes that positioning explicit. Hy3 preview is framed around three priorities: systematic capability, authentic evaluation and cost-effectiveness. That is a different message from the old assumption that the strongest model is simply the biggest one. Tencent is instead arguing for a model that can hold up across reasoning, coding, conversation and tool use while remaining practical enough to serve inside products.
Tencent says the preview is part of a broader rebuild of the Hunyuan model family, giving the company a way to test real-world feedback through open-source release while continuing to expand pre-training and reinforcement learning behind the scenes. That helps frame HY3 Preview not just as a smaller flagship, but as a model being developed around practical adoption as much as raw capability.
Its benchmark choices reinforce that point. Tencent highlights performance not just on broad tests, but on more specific evaluations tied to coding and agent use, including SWE-Bench Verified, Terminal-Bench 2.0, BrowseComp, WideSearch, FrontierScience-Olympiad and IMOAnswerBench. Those are less about spectacle than about demonstrating whether a model can support complicated, multi-step tasks.
Hy3 preview is also already being pushed into products such as Yuanbao, CodeBuddy, WorkBuddy, ima, Tencent Docs and Peacekeeper Elite, which gives Tencent a way to validate the model beyond leaderboard comparisons. That is important because once models reach a certain level of capability, practical adoption is shaped less by scale alone than by whether a model is fast enough, cheap enough and stable enough to be used at volume.
In that sense, the smaller size of Hy3 preview does not look like a limitation. It looks like the product strategy.
Tencent is also opening a two-week free token window, a move that may help it demonstrate the model’s lower-cost, deployment-friendly positioning in practice.
For those interested in exploring Hy3 preview, access is available here (free access for the first two weeks) :
https://openrouter.ai/tencent/hy3-preview:free
Follow on X: @TencentHunyuan
