What is Hy4 flash?

Hy4 flash is Tencent's most cost-effective Hy model, a natively multimodal mixture-of-experts system with a 1 million token context window. The official research page lists 224 billion total parameters, including 64 billion Engram parameters, and 12 billion active per token; the official Hugging Face model card summarizes 220 billion total and the same 12 billion active figure. API and open-weight access are both official.

What Hy4 flash is for

Tencent introduced Hy4 flash on 16 November 2026 as its most cost-effective Hy model to date, aimed at everyday productivity rather than the highest intelligence ceiling. The official research page recommends it for tasks such as deep search, document processing, slide creation, simple frontend demos, and computer use, and it still recommends Hy4 for complex workloads such as long-horizon software engineering.

The company says Hy4 flash is natively multimodal with a 1 million token context window. In Tencent's WorkIndex, an internal suite covering slides, Markdown co-writing, document editing, spreadsheet repair, audio and video processing, and visual understanding, it reports that average cost per task is 1/13 of Hy4 preview, 1/20 of Hy4, and 1/8 of DeepSeek-V4.1-Flash-1017 at Max-tier peak-hour pricing. In a blind evaluation with 85 internal WorkBuddy users across 259 tasks, Tencent says user satisfaction was close to Claude Haiku 5.5, with the shortest end-to-end time and a lower cost per task than the other models it evaluated. Those figures are Tencent's reported results.

On architecture, the official model card describes a 42-layer mixture-of-experts backbone, hybrid sparse and sliding-window attention, LatentMoE, and Engram n-gram embeddings, plus a DFly2 drafter for speculative decoding. The research page and the model card disagree slightly on the headline parameter total, as noted below, but both describe 12 billion active parameters per token in their summary sentences.

A first path to try it

The official product path on the research page is to use Hy4 flash in WorkBuddy, Yuanbao, or another listed Tencent product, or to call the API on Tencent Cloud TokenHub. Through 30 November 2026, Tencent says WorkBuddy access is free and TokenHub API prices are 50% off. Start with a short everyday task the company names, such as editing a document or drafting a slide, and compare the result, the turnaround time, and the billed tokens against what you would expect from Hy4.

If you want to run weights yourself, Tencent has now published Hy4 flash, Hy4 flash-FP8, and the pretrained Hy4 flash-base as open-weight models on Hugging Face (tencent/Hy4-flash), ModelScope, and GitCode, with a GitHub repository at Tencent-Hunyuan/Hy4-flash. The official card documents Transformers, vLLM, and SGLang quickstarts. Hy4 flash-base is a pretrained base model, not an instruction-tuned chat model.

This is a suggested official path, not a recorded test. Confirm the license (Tencent Hy License) before you deploy, and treat OpenRouter as not yet listed as live on the official research page, which only says Hy4 flash will launch there soon.

Pricing, access, and limits

The official research page lists API prices per million tokens of $0.003 cached input, $0.070 input, and $0.281 output for Hy4 flash, compared with $0.042, $0.834, and $2.501 for Hy4. Those are the list rates on the page checked on 18 November 2026. The November promotion, if still offered when you visit, is free use in WorkBuddy and 50% off API pricing via Tencent Cloud TokenHub through 30 November 2026.

The official research page describes 224 billion total parameters, including 64 billion Engram parameters, with 12 billion active per token. The official Hugging Face model card uses 220 billion total parameters, including 64 billion Engram parameters, and 12 billion active per token, and its property table lists 155.5 billion total parameters excluding Engram, ViT, and DFly2, 63.5 billion Engram parameters, 0.43 billion ViT parameters, 0.85 billion DFly2 parameters, and 11.3 billion text plus 0.43 billion vision parameters activated per token. Both sources are official; this profile does not treat one as a silent correction of the other.

Context length is 1 million tokens in both official sources (1,048,576 tokens in the card). Tencent says it will also release open weights for Hy4 soon. Confirm live TokenHub prices and the promotion end date before you budget around them.

What stands out

  • Official 1 million token context window (1,048,576 tokens on the model card) with native text and image input and text output
  • Official API prices, per million tokens, of $0.003 cached input, $0.070 input, and $0.281 output
  • Official open weights for Hy4 flash, Hy4 flash-FP8, and Hy4 flash-base on Hugging Face, ModelScope, and GitCode, under the Tencent Hy License
  • Official availability on Tencent Cloud TokenHub, WorkBuddy, Yuanbao, WeCom, QQ, ima, and Marvis, with a promotion through 30 November 2026
  • Official claimed decoding speed past 150 tokens/s, and WorkIndex cost-per-task figures versus Hy4 preview, Hy4, and DeepSeek-V4.1-Flash-1017

Before you choose

  • The official research page and the official Hugging Face model card use slightly different total parameter totals (224B vs 220B); this profile reports both and does not reconcile them
  • Tencent still recommends Hy4, not Hy4 flash, for complex long-horizon software engineering
  • The WorkBuddy free access and 50% TokenHub API discount end on 30 November 2026 unless Tencent extends them
  • This profile is not a hands-on test of WorkBuddy, TokenHub, or local serving, and OpenRouter is described on the research page only as a forthcoming launch

Sources & further reading

Based on the official documentation below. Features and access can change; check the provider before making a decision.

  1. Official Tencent Hy research page: Introducing Hy4 flash, dated 16 November 2026, rendered 18 November 2026 (opens in a new tab)
  2. Official Hugging Face model card for tencent/Hy4-flash, retrieved 18 November 2026 (opens in a new tab)
  3. Official GitHub repository Tencent-Hunyuan/Hy4-flash, checked 18 November 2026 (opens in a new tab)
Something out of date or incorrect?Report a correction

Was this page useful?

Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.

Share this page

Related reading