Skip to content
NewTry RouterPlus with free creditsTry playground →
← All posts · News

Hy4 flash open weights released under the Tencent Hy License

Instruct, FP8 and base checkpoints are on Hugging Face, with a separate license for large API sellers.

Tencent published open weights for Hy4 flash on November 17, 2026, on Hugging Face, ModelScope and GitCode. The release has three checkpoints: the instruct model, an FP8 build and the pre-trained base model. Tencent launched Hy4 flash as an API model on November 16 and said then that weights would follow (our launch coverage).

What changed

  • Repositories: tencent/Hy4-flash, tencent/Hy4-flash-FP8 and tencent/Hy4-flash-base. Tencent also links a GGUF build at AngelSlim/Hy4-flash-GGUF, which it describes as for end-to-end testing.
  • The model card lists 220B total parameters, including 64B Engram (n-gram embedding) parameters, with about 12B active per token. Tencent's November 16 launch post gave 224B.
  • Mixture of experts with 16 of 512 routed experts plus one shared expert per token. Context is 1,048,576 tokens, with text and image input and text output.
  • A built-in 0.85B drafter, DFly2, handles speculative decoding. reasoning_effort takes high (default, thinking on) or no_think.
  • Tencent recommends vLLM or SGLang and ships prebuilt images: vllm/vllm-openai:hy4-flash and lmsysorg/sglang:hy4-flash. Both model-card examples serve the FP8 checkpoint across 8 GPUs with tensor parallelism.

License

The weights use the Tencent Hy License 1.0, which is derived from Apache 2.0 with added commercial terms:

  • Anyone offering the model to third parties as a general-purpose API or fine-tuning service (MaaS) needs a separate written license from Tencent if their worldwide revenue is above $50 million in any 12-month period.
  • The same applies before using it in a product with more than 100 million monthly active users.
  • Internal use is exempt, as long as the model and its outputs are not made available to third parties.

Hosted price

Tencent lists Hy4 flash on its TokenHub API at $0.070 input, $0.003 cached input and $0.281 output per million tokens, with 50% off through November 30, per its November 16 launch post. The full Hy4 weights have not been released yet (Hy4 launch coverage).

Hy4 flash is not in the RouterPlus catalog as of November 18, 2026.

Sources

  1. Hugging Face: tencent/Hy4-flash model card (November 17, 2026)
  2. Hugging Face: tencent/Hy4-flash-FP8
  3. Hugging Face: tencent/Hy4-flash-base
  4. Tencent Hy License, Version 1.0
  5. Tencent Hy: Introducing Hy4 flash (November 16, 2026)
  6. RouterPlus model catalog (November 18, 2026)