Beating AI News Flash: Tencent Hunyuan officially releases its first model named Flash, Hy4 flash, forming two tiers within the Hy4 series: a flagship and a low-cost Flash.
Hy4 flash has 224B total parameters, of which 64B are Engram memory parameters, with 12B activated per inference. Compared to Hy4's 770B total parameters and 49B activated parameters, the scale is significantly reduced, but it still supports multimodality and a 1 million Token context.
Engram is a conditional memory technology previously proposed by DeepSeek. It stores some common text patterns in a quickly queryable memory table, allowing the model to obtain information by table lookup and reducing repeated computation.
Hy4 flash costs only RMB 0.5 per million Tokens for input, RMB 2 for output, and RMB 0.02 for cache hits. Compared with Hy4's API list prices, these are lower by 12x, 9x, and 15x respectively. However, Hy4 currently also has a limited-time 50% discount, so the actual price gap will be smaller than the list-price gap.
In Tencent's official tests, the PPT collaborative editing score dropped from 78.9 for Hy4 to 74.0, but the cost per task fell from RMB 1.714 to RMB 0.105, about one-sixteenth of the original. The automated task score also only dropped from 56.2 to 52.2, while the cost per task fell from RMB 0.908 to RMB 0.038.
The gap is larger on complex tasks. In the Terminal-Bench 4.0 terminal operation test, Hy4 flash scored only 19.2 points, while the flagship Hy4 scored 44.4 points. The ProgramBench coding test also dropped from 30.0 points to 16.0 points.
