Tencent released Hy4 flash on 16 November 2026, adding a smaller model to its Hy family for work such as document editing and slide creation. In its announcement, the company said Hy4 flash has 224 billion parameters in total but uses 12 billion during each forward pass, an arrangement intended to reduce the computing needed to run it. Tencent lists API input at approximately US$0.070 per million tokens.
Key points
- Hy4 flash uses 12 billion of its 224 billion parameters per forward pass, according to Tencent.
- On Tencent’s internal WorkIndex benchmark, the company reports a per-task cost one-twentieth that of Hy4.
- Tencent lists output at approximately US$0.281 and cached input at approximately US$0.003 per million tokens.
- A blind test involved 85 internal WorkBuddy users assessing 259 tasks across 27 categories.
Hy4 flash activates 12 billion parameters
A parameter is a stored value the model uses to process a request. Hy4 flash has far more of them in total than it uses at any one point: its sparse design selects part of the model for each forward pass, rather than putting every parameter to work each time. The distinction resembles a reference collection from which only the relevant volumes are opened for a particular question, although the whole collection still has to be stored. Tencent counts 64 billion Engram parameters within the model’s 224 billion total.
Tencent says the model accepts multimodal input and supports a context window of up to 1 million tokens, allowing a request to include a long sequence of material. The company also reports that its KV cache, which keeps information from earlier parts of a request available as the model generates more output, is one-sixth the size of Hy4’s and one-24th the size of Hy3’s. That cache is another part of the memory needed to serve a model, separate from its stored parameters.
Editing a document could draw on a long stretch of text within that context window, with only the selected part of the model active as it processes the request. Repeated use of the same input could also draw on the cached-input price Tencent lists.
Tencent says Hy4 flash combines hybrid attention, LatentMoE and Engram features with work done after training to shorten responses. In a release distributed on 17 November, the company put the reduction in average end-to-end task output tokens from reinforcement learning at 50 percent.
The model follows Hy4, a 770-billion-parameter release on 4 November that AI Affairs reported included visual reasoning. AI Affairs subsequently reported a tenfold expansion of Hy4 inference capacity after WorkBuddy users encountered quotas. Tencent recommends Hy4 flash for routine work, while directing more complex tasks, including long-horizon software engineering, to Hy4.
WorkIndex puts Hy4 flash at one-twentieth of Hy4’s task cost
Tencent built WorkIndex around tasks typical of WorkBuddy, its workplace assistant. The company says the benchmark covers creating and editing slides, editing documents, repairing spreadsheets, processing audio and video, and understanding visual material. It evaluates deliverable quality, the length of a single response and the time taken to finish a task, rather than relying solely on scores from public tests.
On that internal benchmark, Tencent reports that Hy4 flash costs one-twentieth as much per task as Hy4 and one-thirteenth as much as Hy4 Preview, which has 770 billion total parameters and 49 billion active parameters. The company puts its per-task cost at one-eighth that of DeepSeek-V4.1-Flash-1017 using the Max tier’s peak-hour pricing. These are comparisons of the costs Tencent measured for WorkIndex tasks under the stated pricing conditions.
For a separate blind evaluation, Tencent says 85 of its internal WorkBuddy users assessed outputs across 259 tasks in 27 categories. The company reports satisfaction close to Claude Haiku 5.5, alongside the lowest per-task cost and shortest completion time among the models it tested. The users, task selection and satisfaction finding described by Tencent are specific to that internal exercise.
TokenHub offers half-price access through 30 November
For developers building on the model, Tencent quotes about US$0.003 for a million cached input tokens, about US$0.281 for a million output tokens and about US$0.070 for a million new input tokens. Those separate charges make the mix of new input, reused input and generated output part of the cost of a task, alongside the model’s computing requirements.
Hy4 flash is available through Tencent Cloud TokenHub and WorkBuddy, as well as Yuanbao, WeCom, QQ, ima and Marvis, Tencent says. WorkBuddy access is free through 30 November, and TokenHub API access is half the listed price during that period, TechNode reported. Tencent also plans to release the model’s weights and make it available on OpenRouter.