// Yapay Zeka

Tencent announced the Hy4 Flash multimodal model

Tencent introduced the Hy4 Flash model with a one-million-token context window. Compared to the full Hy4, the model cuts input and output token prices by about 92% and 89%, respectively.

Tencent Hy introduced the Hy4 Flash model on November 16, 2026. The model has a one-million-token context window and is multimodal by default. But the real standout is the price: Hy4 Flash comes in much cheaper than the full Hy4 version and targets everyday productivity tasks.

Token pricing has dropped to about one-tenth

According to the Tencent Cloud pricing list, Hy4 Flash costs $0.070 per million input tokens and $0.281 per million output tokens. For the full Hy4, those figures are $0.834 and $2.501 respectively. Cached input is also $0.003 per million tokens for Hy4 Flash, compared with $0.042 for the full Hy4. In other words, input pricing is down by about 92% and output pricing by about 89%.

Optimizations in the architecture and service layer

The model contains 224 billion parameters in total, 64 billion of which are Engram parameters. But only 12 billion parameters are active per token. Tencent says the hybrid-attention, LatentMoE, and Engram design reduce the key-value cache: cache size drops to one twenty-fourth of Hy3 and one-sixth of Hy4. The model was trained from Hy4 using self-distillation, and reinforcement learning also cuts the average end-to-end output token count by 50%.

On the service side, Tencent says that thanks to megakernel, four-bit quantization, and DFly, decoding speed exceeds 150 tokens per second. That translates into lower-cost deployment at scale.

WorkIndex measures office tasks

Instead of focusing on public benchmark scores, Tencent has built a metric centered on WorkBuddy tasks. WorkIndex measures output quality, single-turn output length, and task completion time. Its eight evaluation sets cover presentation creation and editing, human-AI collaborative presentation writing, collaborative Markdown writing, document editing, table repair, audio/video processing, and visual understanding tasks.

According to the company, Hy4 Flash performs slightly better overall than the Hy4 preview on WorkIndex. The average cost per task drops to one-thirteenth of the Hy4 preview and one-twentieth of the full Hy4. In addition, in a blind evaluation involving 85 WorkBuddy users across 27 categories and 259 tasks, user satisfaction came out close to Claude Haiku 5.5; Tencent says the model stands out in completion time and cost per task.

Where it can be used

Hy4 Flash is currently available through Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with OpenRouter support also planned. WorkBuddy access is free until November 30, 2026, and TokenHub API usage is 50% off until noon Beijing time on the same day. The API identity for developers is hy4-flash. The documentation lists support for visual understanding, reasoning, structured output, function calling, and caching; maximum input is 960,000 tokens and output is 64,000 tokens. The API supports the OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages formats.

Tencent recommends Hy4 Flash for deep search, document processing, presentation prep, simple UI experiments, and computer use. For long-running, complex software engineering work, it points to the full Hy4 instead. The company also announced an open-weight Hy4 Flash version and plans for frequent updates. That turns the Hy4 family into a two-tier structure: Flash is positioned for high-volume productivity and agent tasks, while the full Hy4 is aimed at harder projects. How well Tencent’s internally measured cost and speed results will hold up under independent workloads, however, remains an open question.

Source: Unite.AI

// JOIN THE CONVERSATION

Share your experience

Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.

// KEEP READING

One more note.

All posts ↗︎