Tencent's Hy4-flash Weights Carry a License That Requires a Separate Deal Above $50 Million in Revenue
A.I. / news
Tencent's Hy4-flash Weights Carry a License That Requires a Separate Deal Above $50 Million in Revenue
The 12B-active model is ungated on Hugging Face, but it ships under the Tencent Hy License, not Apache 2.0.

Tencent's Hy4-flash weights are on Hugging Face, ungated and downloadable, under a licence that is not Apache 2.0. The Tencent Hy License Version 1.0 requires a separate written licence from any operator of a model-as-a-service business with more than US$50 million in worldwide gross revenue, according to the LICENSE file in the repository.
The repository's last modification was timestamped Nov. 18 at 04:24 UTC. It showed 159 downloads and 158 likes when read on Nov. 18, and its licence tag reads "other".
What the Hy4-flash model card says
Tencent's model card describes a mixture-of-experts model that accepts text and images and writes text. It has 12B active parameters per token, with each token routed to 16 of 512 experts plus one shared expert, over 42 backbone layers. The context window is 1,048,576 tokens.
| Spec | Hy4-flash card |
|---|---|
| Total parameters | 220B, including 64B Engram parameters |
| Active parameters | 12B per token |
| Context window | 1,048,576 tokens |
| Weights | Instruct, FP8 and base versions |
The card does not give a parameter count that agrees with every outside account. Explainx's launch write-up lists 224B in one table and 220B in an update, and says Tencent released Hy4 Flash on Nov. 16. The card also gives no training data description beyond noting that the base model saw a large amount of synthetic instruction data in mid-training.

The licence terms that decide who can use it
The licence is derived from Apache 2.0 and says so, adding commercial terms in Section 10. A MaaS business means one that gives third parties remote access to inference or fine-tuning, such as through an API. It needs a separate licence in two cases: worldwide gross revenue above US$50 million in any 12-month period, whether or not the revenue comes from the model, or use in a product with more than 100 million monthly active users.
Internal use is exempt if neither the model nor its outputs reach third parties. End-user products with capabilities confined to specific features are outside the MaaS definition. Licence requests go to an address Tencent lists in the file, hy_partners@tencent.com.
A startup serving the model through an API under US$50 million in revenue falls outside the clause. A large cloud provider reselling it does not.
Hardware and benchmark claims
The card lists no minimum hardware. Its vLLM and SGLang examples use 8-way tensor parallelism with FP8 weights. It claims decoding above 150 tokens per second with optimisations, without naming the hardware.
The benchmark figures are Tencent's own. On its internal WorkIndex suite, Hy4-flash scores 42.1 against 39.7 for Hy4 preview, at about one-thirteenth the cost per task. In a blind evaluation by 85 internal users over 259 tasks, satisfaction was 2.79 out of 4 against 2.81 for Claude Haiku 5.5.
- Claude Haiku 5.52.81 points
- Hy4-flash2.79 points
Source: Tencent Hy4-flash model card, vendor-run, accessed 2026-11-18
Explainx, which republished Tencent's figures, adds that Hy4-flash scores 19.2 on Terminal-Bench 4.0 against 44.4 for the full Hy4, and calls the cost claim closer to 10 times than Tencent's one-twentieth in its own toy calculation. Tencent is running 50% off its TokenHub API pricing through Nov. 30, 2026.
How it compares to other open releases
Alibaba's Qwen3.8-27B is a widely downloaded open model; see its 6.9 million downloads and Artificial Analysis's rating. Thinking Machines' Inkling 2 Small went out under Apache 2.0, which carries no revenue clause.
No independent evaluation of Hy4-flash appeared in the sources read for this piece. The next check is whether Artificial Analysis or Vals lists it, and whether Tencent's licence-request address produces published terms for larger operators.
Sources
More in A.I.
- 01Gemini 4 Carbon Takes 36 Minutes a Task; Kernel Counts 7x the Steps of AstraArtificial Analysis puts Carbon at 36.0 minutes per Intelligence Index task against a 12.4-minute median, and a Kernel study traces most of the gap to extra steps.
- 02Researcher Says Claude Fable 5.2 Learned Protein Sequences; Test Shows a 45-Point GapXinming Tu's probe of 668 obscure bacterial proteins finds the newest Claude models predict real sequences far better than shuffled ones. GPT, Gemini and open-weight models do not.
- 03OpenAI's GPT-6.2 Astra Lists at $8 per Million Input Tokens, 20% Under Its Predecessor, While Coding Scores Stay FlatArtificial Analysis ranks Astra second at 62 behind Claude Fable 5.2, and Sol's coding-agent score fell 3 points.
- 04Google Put Gemini 4 Hydrogen Behind Search for Everyone on Nov. 12, and Published No Click FiguresThe one randomised study of AI Overviews found a 39.8% drop in outbound clicks on early-2026 Search.