Tencent releases Hunyuan Hy4flash: 1/8 the per-task cost of DeepSeek Flash
AGI Hunt · wechat · 2026-11-16
Tencent's Hunyuan team launched Hy4flash, billed as its most cost-efficient model yet: 224B total parameters (including 64B Engram params) with only 12B activated per token, native multimodality, and a 1M-token context window.
Pricing: $0.070/M input tokens, $0.281/M output, $0.003 cache hits — roughly 1/9 to 1/14 of Hy4's rates; 50% off on Tencent Cloud TokenHub until Nov 30.
Benchmarks & cost
- Scores 42.1 on the internal WorkIndex (Hy4 preview: 39.7, Hy4: 51.4) at 0.2 RMB per task — 1/20 of Hy4 and 1/8 of DeepSeek-V4.1-Flash (peak pricing)
- Blind tests by 85 internal users: 2.79/4 satisfaction, essentially tied with Claude Haiku 5.5 (2.81), with the shortest median end-to-end time (6.7 min)
- Office tasks nearly match Hy4 (MdCoEditBench 56.9 vs 57.7), but coding lags far behind: Terminal-Bench 4.0 only 19.2 vs Hy4's 44.4
Architecture & rollout: mixed attention, LatentMoE and Engram for pretraining (KV cache 1/6 of Hy4); self-distillation from Hy4 plus RL that cut average output length by 50%; inference via megakernel and 4-bit quantization at 150+ tokens/s. Live across Tencent Cloud, WorkBuddy, Yuanbao, WeCom, QQ and more, with open-source plans. Tencent recommends it for everyday tasks while keeping Hy4 for long-horizon software engineering.
More from Models
- Dev praises Gemini 4 carbon as his favorite model despite 20-minute waits — willcb · 2026-11-17
- Qwen 3.8 27B vs Claude Fable 5.2: a $0 local model (65 vs 20 on AA Index) and which jobs it can take over — alex_verem · 2026-11-17
- Blogger finds Astra skips steps on research tasks while carbon dutifully finishes them — lxfater · 2026-11-17
- Reuters exclusive: China taps US AI models via distillation to train homegrown rivals — ns123abc · 2026-11-17
- IFM releases K2-Horizon: six fully open LLMs up to 375B MoE, trained on ~10T synthetic tokens — IFM · 2026-11-17
- Optimization researcher: Astra proved with one prompt what we couldn't — bremen79 · 2026-11-17