Tencent releases Hy4flash: 224B MoE with 12B active, 1/20 the cost of Hy4
腾讯混元 · wechat · 2026-11-16
Tencent's Hunyuan team launched Hy4flash, its most cost-efficient model yet: 224B total parameters (including 64B Engram params), 12B active, multimodal, 1M context, with API pricing far below Hy4.
- Pretraining adds hybrid attention, LatentMoE and Engram; KV cache is 1/24 of Hy3's and 1/6 of Hy4's, validated via scaling-ladder experiments.
- Post-training self-distills from Hy4, and RL cuts end-to-end output length by 50%.
- Inference combines megakernel, 4-bit quantization and dflash for 150+ tps decoding at low cost.
On Tencent's internal WorkIndex benchmark (8 internal task suites) plus blind tests with 85 internal users across 259 daily tasks, Hy4flash slightly beats Hy4 preview overall, costs 1/13 of Hy4 preview and 1/8 of DeepSeek-V4.1-Flash per task, with satisfaction near Claude Haiku 5.5. It's live in Yuanbao, WeCom, ima and others, coming to OpenRouter, and will be open-sourced.
More from Models
- HY4 Flash API launches: ~0.2 RMB per task, beats GPT-6 Luna, open source soon — Xianbao_QIAN · 2026-11-16
- Google researcher: open source is becoming a PR play as base model releases dwindle — Guodzh · 2026-11-16
- Model takes first place and $7,500 on ValsAI's PokerBench track — ypatil125 · 2026-11-16
- Dev builds custom standalone Artificial Analysis alternative with Pareto frontiers and dark mode — teortaxesTex · 2026-11-16
- Gemini leaks its own internal instruction "do not use python" into a Google AI Overview — bclavie · 2026-11-16
- Qwen4-27B incoming: insider says Alibaba's small model lands in days, near-zero inference cost — bindureddy · 2026-11-16