WorkBuddy runs many model calls per task, so a model this cheap lets Tencent offer agent work at scale on less of the scarce high-end computing that capped Hy4.
Tencent released Hy4 flash on November 16, a smaller and cheaper member of its Hunyuan model family built for routine office work. On Tencent’s own task benchmark, the model completed a typical job at about a twentieth of the cost of the full Hy4 model released 12 days earlier. Science and Technology Daily reported the launch the same evening, and Tencent’s English release calls it the most cost-effective Hy model to date.
The model has 224 billion parameters, 64 billion of them in a component Tencent calls Engram, but only 12 billion are active for each token it processes, which keeps running costs low. It accepts multimodal input and handles up to 1 million tokens of context. WorkBuddy, Tencent’s office agent, offers it free until November 30, and Tencent Cloud’s TokenHub sells API access at half price over the same period.
All of the performance figures come from Tencent’s internal tests, and no outside benchmark has yet checked them. The company says it plans to release the weights but has not given a date.
Prices fall by an order of magnitude against Hy4
Tencent lists Hy4 flash at RMB0.5 per million input tokens and RMB2 per million output tokens, against RMB6 and RMB18 for the full Hy4. Reused text from a cache costs RMB0.02 per million tokens, down from RMB0.3. Those Hy4 prices come from Tencent’s Hy4 launch page of November 4. By EastFrontier’s calculation, the new model costs a twelfth as much for input, a ninth for output and a fifteenth for cache hits. In dollars, Tencent quotes about US$0.070 and US$0.281 per million input and output tokens.
The per-task figure goes further than the price list because the model also writes less. According to IT Home, reinforcement learning cut the average length of a task’s output by half, and the working memory the model holds during a conversation is a sixth the size of Hy4’s and a twenty-fourth of Hy3’s. Decoding runs above 150 tokens a second.
On WorkIndex, an internal benchmark built from typical WorkBuddy jobs such as making and editing slides, repairing spreadsheets and processing audio and video, Tencent says Hy4 flash slightly outscored the Hy4 preview, a 770-billion-parameter model. Its cost per task was a thirteenth of the preview’s, a twentieth of Hy4’s and an eighth of DeepSeek-V4.1-Flash-1017 at that model’s top setting and peak-hour prices.
A blind test puts it close to Claude Haiku 5.5, by Tencent’s account
Tencent also ran a blind comparison with 85 of its own WorkBuddy users across 259 tasks in 27 categories. It reports that their satisfaction with Hy4 flash came close to Claude Haiku 5.5, and that Hy4 flash had the lowest cost per task and the shortest completion time of the models tested.
The company is clear about the limits. It recommends Hy4 flash for most daily work, including deep search, document processing, slide generation, simple front-end pages and computer use, in which the model operates desktop software. For long, complex software engineering it still points users to Hy4.
The model is live on TokenHub, WorkBuddy, the Yuanbao assistant, WeCom, QQ, the ima knowledge app and Marvis, with OpenRouter listed as coming soon. In WorkBuddy, Landian reports, the default context window is 300,000 tokens, adjustable to 600,000 or 960,000, and larger windows use up the daily free allowance faster.
The end of the free period will test demand for a cheaper default
Tencent has reason to push work onto a lighter model. Hy4’s launch outran the computing set aside for it, and Tencent expanded the Hy4 cluster tenfold after WorkBuddy users hit their quotas. Hy4’s free WorkBuddy quota ran from November 4 to November 15. Hy4 remains at half its list price on TokenHub through December 31.
TechNode summarized the 95% saving as the headline result. Three markers will show whether it holds outside Tencent’s labs: usage after the free period ends on November 30; the model’s arrival on OpenRouter, where developers compare prices openly; and the release of weights that would let outside testers rerun Tencent’s tests.