Tencent's Hy4 flash significantly lowers AI API costs
Tencent introduced Hy4 flash, a lighter version of its multimodal Hy4 model, with a context window of one million tokens and significantly lower API prices. The goal is clear: target high-volume productivity tasks without sacrificing quality.
Tencent launched Hy4 flash on November 16, 2026, a multimodal version of its Hy4 model designed for everyday productivity tasks. With a context window of one million tokens and API pricing much lower than Hy4, the group aims to make AI usage more sustainable for large-scale repetitive tasks. The bet pays off, as inference cost remains one of the real battlegrounds for companies. And, incidentally, bills rarely have a sense of humor.
Broken prices, but with a precise positioning
Tencent Cloud's pricing grid lists Hy4 flash at $0.070 per million input tokens and $0.281 per output million tokens. Compared to Hy4, charged at $0.834 per input and $2.501 per output on the same basis. On the listed prices alone, this represents about 92% less on input and 89% less on output.
The commercial message is clear: Hy4 flash targets high-volume uses, while Hy4 retains the role of the more ambitious model for complex tasks. Tencent recommends it for in-depth research, document processing, slide creation, simple front-end demonstrations, and computer use by agents. For long-term software projects, the group continues to favor the full version, considered more capable.
A streamlined architecture to keep up the pace
Tencent says it built Hy4 flash on a hybrid architecture combining latent attention, LatentMoE, and Engram. In short, the model reduces the size of the key-value cache, that is, the temporary memory that reuses calculations on the already seen context. According to the publisher, this cache is reduced to one twenty-fourth of Hy3's and one sixth of Hy4's.
The model totals 224 billion parameters, including 64 billion Engram parameters, with 12 billion activated per token. After training, Tencent indicates it used self-distillation from Hy4, then reinforcement that reduces the average number of tokens produced end-to-end by 50%. On the execution side, megakernels, four-bit quantization, and DFly push decoding speed beyond 150 tokens per second according to the company. This remains an internal promise at this stage, but the trajectory is clear: produce faster, cheaper, without driving up the bill at each interaction.
WorkIndex, or how to judge a model on real office tasks
Rather than highlighting traditional public benchmarks, Tencent built WorkIndex around real office tasks. The evaluation measures deliverable quality, single-pass output length, and processing time. The eight test sets cover slide creation and editing, co-writing with humans, document editing, spreadsheet repair, audio/video processing, and visual understanding.
On WorkIndex, Tencent claims that Hy4 flash performs slightly better than Hy4 preview overall, while bringing the average cost per task down to one thirteenth of Hy4 preview, one twentieth of Hy4, and one eighth of DeepSeek-V4.1-Flash-1017 at Max pricing during peak hours. Another blind evaluation involved 85 internal WorkBuddy users on 259 tasks across 27 categories. Announced result: satisfaction close to Claude Haiku 5.5, with the shortest task completion time and the lowest cost among tested models.
A two-tier strategy for Tencent products
Hy4 flash is already available in Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with an arrival planned on OpenRouter. Access to WorkBuddy is free until November 30, 2026, while usage via TokenHub benefits from a 50% discount until noon Beijing time the same day. For developers, the API identifier is hy4-flash, supporting OpenAI Chat Completions alignment, OpenAI Responses, and Anthropic Messages, as well as caching functions.
This release follows the November 4, 2026 announcement on Hy4, which introduced visual understanding and computer use capabilities. Tencent is also preparing an open-weight version of Hy4 flash and promises frequent updates. The group thus establishes a fairly clear duo: one model to handle large workloads, another for more difficult cases. It remains to be seen whether the cost and speed gains observed in the Tencent environment will hold up outside its own services, which is often when the fine print really matters.
Key points
- Hy4 flash was released on November 16, 2026.
- API pricing: $0.070 input, $0.281 output.
- Hy4 flash targets high-volume productivity tasks.
- WorkIndex covers eight office test sets.
- 85 internal users participated in the blind evaluation.
- Tencent is preparing an open-weight version of Hy4 flash.
By the numbers
- 1 million tokens — announced context window for Hy4 flash.
- 224 billion parameters — total model size, according to Tencent.
- 150 tokens per second — decoding speed announced by Tencent.
- 50% — average reduction in output tokens after reinforcement, according to Tencent.
- 259 tasks — volume of the internal WorkBuddy evaluation cited by Tencent.