律动BlockBeats
律动BlockBeats|11月 16, 2026 06:14
[Tencent Launches First Flash-Based Model: 224B Total Parameters, 12B Activation, Costs Reduced by Over 20 Times for Some Office Tasks] Beating AI News: Tencent's Hy4 flash, the first model named "Flash," officially launched, creating two tiers within the Hy4 series: a flagship model and a low-cost Flash model. The Hy4 flash has 224B total parameters, of which 64B are Engram memory parameters, with 12B activated per inference. Compared to the Hy4's 770B total parameters and 49B activation parameters, the scale is significantly smaller, yet it still supports multimodal processing and a context of 1 million tokens. Engram is a conditional memory technology previously proposed by DeepSeek. It stores common text patterns in a quickly searchable memory table, allowing the model to retrieve information by looking up the table, reducing redundant computation. The Hy4 flash costs only 0.5 yuan per million token inputs, 2 yuan per output, and 0.02 yuan per cache hit. Compared to Hy4's API list price, the prices are reduced by 12 times, 9 times, and 15 times respectively. However, Hy4 currently has a limited-time 50% discount, so the actual price difference will be less than the list price difference. In Tencent's official tests, the PPT collaborative editing score dropped from Hy4's 78.9 to 74.0, but the cost per task decreased from 1.714 yuan to 0.105 yuan, approximately one-sixteenth of the original. The automation task score also only decreased from 56.2 to 52.2, while the cost per task decreased from 0.908 yuan to 0.038 yuan. The difference is even greater for complex tasks. In the Terminal-Bench 4.0 terminal operation test, Hy4 flash only scored 19.2 points, while the flagship Hy4 scored 44.4 points. The ProgramBench programming test also dropped from 30.0 points to 16.0 points. [Original Link]
+2
Mentioned
Share To

Timeline

HotFlash

APP

X

Telegram

Facebook

Reddit

CopyLink

Hot Reads