Tencent releases Hy4 Flash with AI input priced at $0.07 per million tokens
The model reaches Tencent’s apps and cloud API. Its claimed savings on completed office tasks come from internal tests, while the flagship remains recommended for complex software projects.
With Hy4 Flash, Tencent is adding a separate, lower-cost hosted model for high-volume productivity work while recommending full Hy4 for complex, long-running software engineering. Flash’s API rates are $0.070 per million input tokens and $0.281 per million output tokens—roughly 92% and 89% below full Hy4. The model is available through Tencent services now, but downloadable weights are only planned, so developers seeking to self-host cannot yet use the announced release that way.
01
Tencent says Flash has 224 billion total parameters, with 12 billion active per token, and uses one-sixth Hy4’s cache size.
02
On Tencent’s WorkIndex evaluation, Flash scored slightly above Hy4 Preview overall at one-thirteenth the average task cost; these are company-run results.
03
In a separate internal test, 85 WorkBuddy users completed 259 tasks; Tencent reports satisfaction close to Claude Haiku 5.5 and the lowest task cost among evaluated models.
Everyday AI work now has a cheaper option in Tencent’s lineup. The company released Hy4 Flash on November 16, 2026, with API rates of $0.070 per million input tokens and $0.281 per million output tokens. It targets high-volume productivity work, while Tencent still recommends full Hy4 for complex, long-running software engineering.
The release delivers the lower-cost version Tencent previewed earlier this month. In its November 4 announcement for full Hy4, the company said Flash would arrive soon for users prioritizing cost efficiency. Flash now joins that model as a separate option, rather than replacing it.
Lower prices, less computation per answer
Tencent Cloud’s listed full-Hy4 rates are $0.834 per million input tokens and $2.501 per million output tokens, according to Unite.AI. Against those prices, Flash cuts the input rate by about 92% and the output rate by about 89%. Cached input, which reuses earlier input computations, has a separate, lower rate.
Hy4 Flash’s listed API rates
$0.070Input
The listed input rate is separate from the discounted cached-input rate.
$0.281Output
Generated output is billed at $0.281 for every million tokens.
$0.003Cached input
Cached input costs $0.003 per million, compared with $0.042 for full Hy4.
Tencent reports 224 billion total parameters—the model’s learned settings—including 64 billion Engram parameters. Only 12 billion parameters activate for each token. The company also says its design shrinks the cache storing earlier context computations to one-sixth of Hy4’s size. These are architecture and efficiency claims supplied by Tencent.
Training and serving changes contribute to the efficiency pitch. Tencent says further training using Hy4 as a teacher, followed by reinforcement learning, cut average end-to-end output-token use by 50%. It also reports decoding speeds above 150 tokens per second after serving optimizations. Those figures describe Tencent’s results, rather than a guaranteed speed for every request.
Office deliverables are the test
Tencent’s WorkIndex evaluation measures deliverable quality, single-turn output length and turnaround time on WorkBuddy tasks. Its eight evaluation sets include slide creation and editing, collaborative writing, document editing, spreadsheet repair, audio/video processing and visual understanding. The emphasis is on finished office work rather than a ranking built around public benchmark scores.
On that internal evaluation, Tencent says Flash performed slightly above Hy4 Preview overall, with average task costs one-thirteenth as high. It reports task costs at one-twentieth of full Hy4’s. Those are costs for completing the evaluated assignments, not additional across-the-board reductions in the token rates shown above.
A separate blind evaluation involved 85 internal WorkBuddy users completing 259 tasks across 27 categories. Tencent reports satisfaction close to Claude Haiku 5.5, alongside the shortest completion time and lowest task cost among the evaluated models. The comparison comes from Tencent’s own users and testing, not an independent assessment.
Hosted access now, downloadable weights later
Tencent describes Flash as natively multimodal, with image understanding and a one-million-token context window—the material it can work with in an interaction. Cloud documentation lists maximum input and output limits of 960,000 and 64,000 tokens, respectively, alongside reasoning, structured output, function calling and caching.
Developer access: Tencent Cloud TokenHub uses the identifier hy4-flash and supports OpenAI Chat Completions, OpenAI Responses and Anthropic Messages interfaces. Customers must activate the model service first.
App access: Flash is available through WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis. OpenRouter availability is planned.
Launch offers: WorkBuddy access is free through November 30, 2026. TokenHub usage receives a 50% discount until noon Beijing time that day, covering usage during the promotional period.
Tencent also plans to release Flash’s model weights. For now, the release gives users hosted access through its services; the downloadable version remains a future step.
Sources
tencent.comTencent Launches Hy4, Joining the Ranks of the World's Leading Open-Source AI Models
unite.aiTencent Launches Hy4 Flash as a Lower-Cost Multimodal Model
Reader comments
Newest comments first. Replies stay oldest first.