Reported plans
Tencent's Hy team outlines larger models and 24GB deployment plans
- In a Reddit AMA,
Tencent's Hy team said the next Hy model would be substantially larger and that it is planning smaller models for devices with 24GB of GPU memory.
- The team said the next Hy model would be substantially larger and that it would continue investing in large-scale reinforcement learning, according to a Chinese-language compilation by Zhidx.
- At the smaller end, the team described Flash-Lite as planned work and discussed models aimed at devices with 24GB of GPU memory, while a dense model in the 24-billion-to-31-billion-parameter range was under consideration.
- The team also discussed trade-offs between grouped-query attention and multi-head latent attention, saying MLA offered better model quality at a comparable key-value-cache size but its compute intensity complicated decoding.
- On training, the team said the hardest behavior to improve in post-training was 'overthinking,' and that for Hy4 flash they first distilled from the flagship then used reinforcement learning to shorten the reasoning chain.
Source details
- Latest report
- 21 Nov 2026 06:29 UTC
- First seen here
- 21 Nov 2026 06:29 UTC
- Outlets
- 2