Tencent launches Hunyuan 4flash: 224B-total/12B-active, tasks cost $0.036

智东西 · wechat · 2026-11-16

Tencent released Hunyuan 4flash, a 224B-total/12B-active multimodal model with 1M-token context and 150+ tok/s decoding. Priced at ¥0.5/¥2 per million input/output tokens (1/12 and 1/9 of Hunyuan 4). In an 85-user, 259-task blind test it scored 2.79/4 satisfaction, nearly matching Claude Haiku 5.5 at 1/18th the per-task cost (¥0.26), with the fastest completion time. Cost cuts come from hybrid attention, Latent MoE, 6x KV-cache compression, RL-shortened outputs, and 4-bit quantization. Tencent plans to open-source it.

Related event: Tencent Hunyuan Launches Hy4 flash, Matching Haiku 5.5 in Blind Tests at ~1/18 the Cost(18 posts)→

Original post →

More from Models

Models channel →