Tencent releases Hy4 flash: 1 million output tokens cost about $0.28
Tencent’s Hy4 flash language model costs about $0.07 per million input tokens and about $0.28 per million output tokens. Tencent says the model performs slightly better overall in its own benchmarks than the larger Hy4 preview.
Tencent’s Hunyuan team announced a new language model, built to be cheap and fast for everyday work, in a blog post and a company press release on 16 November. Hy4 flash has 224 billion parameters, 64 billion of which are so-called Engram parameters.
Hy4 flash is a mixture-of-experts (MoE) model: it uses only some of its parameters, in this case 12 billion, to process each token. The model accepts more than just text as input and has a context window of 1 million tokens.
The table shows the API prices that Tencent’s blog gives for Hy4 flash and the larger Hy4 model. Pandaily and InfoQ China give Hy4 flash’s yuan prices as 0.5 for input, 2 for output and 0.02 for cached input.
| Price per 1 million tokens | Hy4 flash | Hy4 |
|---|---|---|
| Cached input | $0.003 | $0.042 |
| Input | $0.070 | $0.834 |
| Output | $0.281 | $2.501 |
Source: Tencent Hy blog, 16 November 2026; Tencent’s press release says the dollar prices are approximate.
The cost of a task depends not only on the price per token but also on how much text the model generates.
How Tencent cut the price
Tencent added hybrid attention, LatentMoE and Engram to the model. Pandaily describes Engram as a parameter module that works like a lookup table. While generating a response, the model stores information about text it has already processed in a KV cache, which the company says is 1/24 the size of Hy3’s cache and 1/6 the size of Hy4’s.
In post-training, Hy4 served as the teacher model, and reinforcement learning halved the average amount of text the model generates per task. During inference, 4-bit quantisation, megakernels and DFly push speed above 150 tokens per second.
The results come from Tencent’s own tests
Tencent built the WorkIndex test from typical tasks in its WorkBuddy service, such as creating slide decks, editing documents, fixing spreadsheets, processing audio and video, and understanding images. Pandaily reports that the test is not public.
Tencent says Hy4 flash performs slightly better overall than Hy4 preview, which has 770 billion parameters, 49 billion of them active. On WorkIndex, Hy4 flash’s average cost per task is 1/13 of Hy4 preview’s, 1/20 of Hy4’s and 1/8 of DeepSeek-V4.1-Flash-1017’s when DeepSeek runs on the Max tier at peak-hour prices, the company said. InfoQ China gives the DeepSeek version as 1027.
In the WorkBuddy team’s own evaluations, average task time fell by 70 per cent, and the cost was about 1/15 of Hy4’s. The team said quality matched Hy4 on everyday tasks but fell slightly behind on long coding tasks. Tencent still recommends Hy4 for long software development tasks.
In a blind test, 85 internal WorkBuddy users at Tencent rated 259 tasks across 27 categories. Tencent said Hy4 flash’s user satisfaction was close to that of Claude Haiku 5.5, its cost per task was clearly lower than that of the other models tested, and its task times were the shortest.
Where to get the model
Hy4 flash is available on services including Tencent Cloud TokenHub, WorkBuddy and Yuanbao, and is coming to OpenRouter. The model is free on WorkBuddy until 30 November, and TokenHub API prices are 50 per cent off until then. Nous Research said on X that the model is free on Nous Portal for two weeks, and OpenClaw said the model is now available in OpenClaw. Tencent says it plans to release Hy4 flash as an open-weight model.
Read more: Models