Tencent Hunyuan says its next model will be significantly larger, while it is also planning models for devices with 24GB of VRAM. The roadmap points in two directions at once: scaling up frontier models and making smaller options for local use.
Tencent AI Watch analysis
What happened
In a Reddit AMA, the Hunyuan team said Hy4-mini is on its development roadmap and that more small models are planned. It also confirmed work on models for devices with 24GB of video memory, alongside continued development for systems with 64GB to 128GB of memory. The team did not give release dates for these plans.
The details come from 36Kr’s account of the AMA. It also reports that Tencent intends to keep investing in large-scale reinforcement learning and to expand the scale of its next-generation Hy model.
Why it matters
A model that can run within 24GB of VRAM could make local inference practical for more developers and operators, without requiring data to be sent to a cloud service. That is a stated direction, not a promise that a particular model is available now: Tencent gave no specifications, performance figures or timetable for the planned device models.
The larger-model plan shows that local options are not a replacement for Tencent’s scaling ambitions. The company is pursuing both, while saying it sees room to improve Hy4 through post-training.
Our read
The useful news is the combination: bigger models at the high end, and planned options for more modest local hardware. But roadmap language is cheap; the test will be whether the smaller models arrive with useful capabilities and clear hardware requirements. For now, developers can treat this as a direction of travel, not a shopping list.
What to watch
- Whether Tencent publishes a release date and specifications for its 24GB VRAM models.
- How Hy4-mini and the next-generation model compare on practical tasks, not just benchmarks.
- Whether the promised local models work with widely used inference tools.
Discussion spark: Would you choose a capable model that runs locally on a 24GB GPU over a stronger cloud model, if the local version required more setup?
Sources and evidence
Independent WittyWires tracker for public updates about Tencent AI. Not affiliated with or endorsed by Tencent AI; this is not an official account.