Deyron Labs graphic: Tencent Hy4 flash open weights, 20 Nov 2026; 220B total parameters, 12B active per token; 1M-token context; Tencent Hy License (custom); Benchmarks are Tencent's own

What happened

Tencent launched Hy4 flash, a model in its Hunyuan family aimed at everyday work tasks, on 16 November 2026 through its apps and cloud API, according to TechNode. The weights are published on Hugging Face under tencent/Hy4-flash, with an FP8 version and a base (not chat-tuned) version. We read the model card on 20 November 2026; we have not run the model, and we could not read Tencent’s own press release.

Key details

From the model card, unless noted. Benchmark and cost figures are Tencent’s own and have not been independently checked.

  • Size: 220B total parameters, including 64B “Engram” parameters, with 12B active per token (the card also lists 155.5B excluding Engram, vision encoder and drafter). TechNode’s articles say about 224B; we use the card’s figure.
  • Inputs and context: text and images; a 1M-token context (1,048,576 tokens). Architecture listed: 42 layers, 512 routed experts per MoE layer with 16 active plus one shared expert.
  • Speed and efficiency (Tencent): decoding “past 150 tokens/s”; KV cache described as 1/24 of Hy3 and 1/6 of Hy4; reinforcement learning said to cut average output tokens by 50%.
  • WorkIndex (Tencent’s internal benchmark): Hy4 flash 42.1, Hy4 preview 39.7, Hy4 51.4. Tencent says the average cost per task is 1/13 of Hy4 preview, 1/20 of Hy4 and 1/8 of DeepSeek-V4.1-Flash-1017 under the conditions it lists.
  • Blind user test (Tencent): 85 internal WorkBuddy users, 259 tasks, 27 categories; satisfaction 2.79 out of 4 for Hy4 flash versus 2.81 for Claude Haiku 5.5.
  • Usage claims (Tencent via TechNode): more than 50 trillion tokens on the first day, and first place on the OpenRouter daily leaderboard on 17 November. We did not verify either.
  • Price (TechNode): 0.5 yuan (about $0.07) per million input tokens and 2 yuan (about $0.30) per million output tokens; we did not confirm this on a Tencent page.
  • License: “Tencent Hy License”, a custom license (Hugging Face tag other). We have not reviewed its terms, so check them before commercial use.

Why it matters

Open weights at this size with a million-token context give developers outside Tencent’s cloud a way to run and fine-tune a model that Tencent positions around low cost per task. Because the license is custom, it is not the same as Apache 2.0 or MIT, and the terms decide what you can do with it.

The headline comparisons (cost per task, user satisfaction versus Claude Haiku 5.5) come from Tencent’s internal tests, so treat them as claims until independent evaluations appear.

Disclosure: Claude, made by Anthropic, was one of the AI tools used to research and draft this article. Claude Haiku 5.5, an Anthropic model, is mentioned only as a comparison reported by Tencent.

Sources

  1. tencent/Hy4-flash model card Primary source , Tencent on Hugging Face (repository created 17 November 2026)
  2. Tencent says Hunyuan Hy4 flash surpasses 50 trillion tokens in first-day usage TechNode, 19 November 2026
  3. Tencent releases Hy4 flash model weights for local deployment TechNode, 19 November 2026