Tencent releases Hy4 flash, low-price tier model shows weaker performance on complex tasks
BIBIBI
AT A GLANCE
Tencent Hunyuan releases Hy4 flash, focusing on lower prices, but complex task test scores are lower than Hy4.
Article
Tencent debuts its first Flash model, approaching Hy4 in PPT revision scores at a 16 times lower cost
Beating AI reported that Tencent Hunyuan released Hy4 flash, its first model named Flash, creating flagship and lower-cost Flash tiers within the Hy4 series.
Hy4 flash has 224B total parameters, including 64B Engram memory parameters, and activates 12B per inference. That is substantially smaller than Hy4, which has 770B total and 49B active parameters, while still offering multimodal support and a context window of 1 million Token.
Engram is a conditional memory technique introduced previously by DeepSeek. It stores some common text patterns in a fast lookup table so the model can retrieve information and reduce repeated computation.
For each million Token, Hy4 flash costs 0.5 yuan for input, 2 yuan for output and 0.02 yuan for cache hits. These are reductions by factors of 12, 9 and 15 versus Hy4’s API list prices, respectively. Hy4 is currently available at half price for a limited time, making the effective gap smaller than the list-price gap.
In Tencent’s tests, the PPT collaborative editing score fell from 78.9 for Hy4 to 74.0, while cost per task fell from 1.714 yuan to 0.105 yuan, about one part in 16 of the original cost. The automation score fell from 56.2 to 52.2, while cost per task fell from 0.908 yuan to 0.038 yuan.
The gap is wider on complex tasks. Hy4 flash scored 19.2 in Terminal-Bench 4.0, compared with 44.4 for flagship Hy4. Its ProgramBench coding score also fell from 30.0 to 16.0.
Reports say Tencent Hunyuan released its first model named Flash, Hy4 flash.
02
Materials state Hy4 flash has a total of 224B parameters, of which 64B are Engram memory parameters, with 12B activated per inference; Hy4 has 770B total parameters and 49B activated.
03
Materials state Hy4 flash supports multimodality and 1000000 Token context.
04
Hy4 flash prices per million Tokens for input, output, and cache hits are 0.5 yuan, 2 yuan, and 0.02 yuan respectively; compared to Hy4 API list prices, these are reduced by 12-fold, 9-fold, and 15-fold respectively.
05
The source says Hy4 is currently available at half price for a limited time, so the effective price gap is smaller than the list-price gap.
06
In Tencent's official tests, the PPT collaborative editing score dropped from Hy4's 78.9 to 74.0, and the cost per task fell from 1.714 yuan to 0.105 yuan.
07
In Tencent's official tests, the automation task score dropped from 56.2 to 52.2, and the cost per task fell from 0.908 yuan to 0.038 yuan.
08
In the Terminal-Bench 4.0 terminal operation test, Hy4 flash scored 19.2, while Hy4 scored 44.4; the ProgramBench programming test score dropped from 30.0 to 16.0.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This news introduces Tencent Hunyuan's Hy4 flash. Its price is lower, and costs in some task tests dropped significantly; but in terminal operation and programming tests, scores are lower than the flagship Hy4. The report also says it supports multimodality and 1000000 Token context.
Why it matters to readers
For those using model APIs, the material shows the trade-off between price and task performance: some tasks may be cheaper, but complex task test scores are significantly lower. Test results come from Tencent officials, and the material does not provide independent evaluations.
If you are choosing a model, you can take low price as one reference, but also consider your own task type. You can first compare quality and cost on the same tasks; the test data in the report cannot guarantee your actual results will be the same.
Risks and unknowns
Parameters, prices, and test data come from the report, and test results are attributed to Tencent officials; they cannot be independently verified based solely on this material.
The material does not specify the specific methods, scoring ranges, or margins of error for each test.
Hy4’s half-price offer affects the actual price gap. The source gives no duration or terms for the offer.
The material does not mention user adoption, actual inference costs, or commercialization performance.
Related Developments
Loading event timeline…
Related concepts
Token: The billing and context length unit in the material.
This term is not in the glossary yet. Browse related concepts in the glossary.
View glossary →Parameters: The numerical scale used for computation in the model; parameter count itself is not equivalent to task performance.
This term is not in the glossary yet. Browse related concepts in the glossary.
View glossary →Activated parameters: The number of parameters activated per inference as stated in the material.
This term is not in the glossary yet. Browse related concepts in the glossary.
View glossary →Engram: The material calls it a conditional memory technology that can place common text patterns into a memory table for quick lookup.
This term is not in the glossary yet. Browse related concepts in the glossary.