Innovation & Tech

Tencent Cuts Hy4 Flash Token Prices Up to 92% in Multimodal Push

Tencent Hy launched Hy4 flash on November 16, 2026, with token prices up to 92% below full Hy4 and a one-million-token context window aimed at high-volume office productivity work.

By Amara Osei

3 min read

Tencent Launches Hy4 Flash as a Lower-Cost Multimodal Model
Tencent Launches Hy4 Flash as a Lower-Cost Multimodal Modelblickpixel / pixabay

What's News

  • Hy4 flash input tokens cost $0.070 per million, about 92% below Hy4's $0.834, per Tencent Cloud's price list.
  • The model has 224B total parameters, activates 12B per token, and carries a one-million-token context window.
  • Tencent reports decoding beyond 150 tokens per second via megakernels, four-bit quantization and DFly.
  • WorkBuddy access is free through November 30, 2026, with a 50% TokenHub API discount until 12:00 Beijing time.
  • Average cost per task is one-twentieth of Hy4 and one-eighth of DeepSeek-V4.1-Flash-1017 at Max-tier pricing, Tencent says.

Tencent Hy has cut input token prices by roughly 92% with Hy4 flash, a natively multimodal model launched on November 16, 2026, that undercuts the full Hy4 model on nearly every cost metric. Input tokens cost $0.070 per million and output tokens $0.281 per million, against $0.834 and $2.501 for Hy4, according to Tencent Cloud's price list. Cached input runs at $0.003 per million tokens, compared with $0.042 for Hy4.

The model carries a one-million-token context window and targets everyday productivity work rather than frontier reasoning. It ships less than two weeks after Tencent introduced full Hy4 on November 4, 2026, which added visual understanding and computer-use capabilities. That earlier announcement flagged a cheaper Flash version as coming soon.

How does the architecture keep costs down?

Hy4 flash contains 224 billion total parameters, including 64 billion Engram parameters, but activates only 12 billion parameters per token. Tencent says its hybrid-attention, LatentMoE and Engram design shrinks the key-value cache — the stored context computations reused during generation — to one twenty-fourth of Hy3's size and one-sixth of Hy4's.

Post-training relies on self-distillation from Hy4, with reinforcement learning cutting average end-to-end output tokens by 50%. On the serving side, Tencent credits megakernels, four-bit quantization and DFly with pushing decoding beyond 150 tokens per second, which the company says enables lower-cost deployment at scale.

What is WorkIndex?

Tencent built its own evaluation, WorkIndex, around typical WorkBuddy tasks instead of public benchmark scores. It measures deliverable quality, single-turn output length and task turnaround time across eight evaluation sets:

  • Slide creation and editing
  • Human-AI slide co-writing
  • Human-AI Markdown co-writing
  • Document editing
  • Spreadsheet repair
  • Audio/video processing
  • Visual understanding

Tencent reports that Hy4 flash scores slightly above Hy4 preview overall on WorkIndex while average cost per task falls to one-thirteenth of Hy4 preview, one-twentieth of Hy4 and one-eighth of DeepSeek-V4.1-Flash-1017 at Max-tier peak-hour pricing.

A separate blind evaluation put 85 internal WorkBuddy users through 259 tasks across 27 categories. Tencent reports user satisfaction close to Claude Haiku 5.5, with the shortest end-to-end completion time and the lowest cost per task among the models evaluated.

Where can developers get it?

Hy4 flash is available through Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima and Marvis, with OpenRouter availability planned. WorkBuddy access is free through November 30, 2026, and TokenHub API usage gets a 50% discount until 12:00 Beijing time that day. The discount covers only usage generated during the promotional period.

The API identifier is hy4-flash. Tencent Cloud's documentation lists image understanding, reasoning, structured output, function calling and caching, with maximum input and output limits of 960,000 and 64,000 tokens respectively. The API supports OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats, and requires customers to activate the model service before making calls.

Tencent recommends Hy4 flash for deep search, document processing, slide creation, simple frontend demonstrations and computer use. Full Hy4 stays the recommendation for complex, long-horizon software engineering because of its higher intelligence ceiling. Tencent also plans an open-weight Hy4 flash release and frequent updates.

The result is a two-tier lineup: Hy4 flash for high-volume productivity and agent work where token use and turnaround time dominate, and full Hy4 for harder software projects. Whether Tencent's internal cost and speed results carry over to independent workloads — and to the planned open-weight release — will determine how far the efficiency gains extend beyond Tencent's own services.

Source: unite.ai

Share this article:

More from Amara Osei

Amara Osei

Show full bio

Senior reporter covering companies, deals and strategy in entrepreneurship, startups and venture funding at Business Bearings.

2,136 articles

Related articles

« Previous articleNext article »