Hy4 open weights on BenchGecko · 770B MoE under the Tencent Hy License
Tencent released Hy4 weights on November 20, 2026 · 770B total and 49B active parameters · 1,048,576 token context · BenchGecko now lists Hy4 and Hy4 flash as open weights.
What changed
Tencent published the weights of Hy4 on November 20, 2026, on Hugging Face, ModelScope and GitCode. Three checkpoints are public: the instruct model Hy4, an FP8 build Hy4-FP8 and the pretrained Hy4 base. Hy4 has been tracked on BenchGecko since its API launch on November 4, 2026. Its row now shows open weights under the Tencent Hy License instead of a proprietary license.
Model facts
- Architecture · mixture of experts · 770B total parameters · 49B active per token
- Layers · 78 · 1 dense layer and 77 MoE layers · 256 routed experts plus 1 shared expert · top 8 routing
- Vision encoder · 1.23B parameters · text, image and video input
- Context window · 1,048,576 tokens
- Speculative decoding drafter · DFly2 · 3.3B parameters
- License · Tencent Hy License, not Apache 2.0
The license is a custom Tencent license. Teams that plan to serve Hy4 commercially should read its terms before deploying.
Price on BenchGecko
BenchGecko tracks 1 provider offer for Hy4: Tencent at $0.42 input and $1.25 output per million tokens. With weights public, more hosts can now serve the model, and BenchGecko will add new provider offers as they appear in its daily catalog sync.
The Hy4 family
Hy4 flash is the lighter tier, with 224B total parameters and 12B active per token. Its weights went public on November 18, 2026, and its row also moves to open weights under the Tencent Hy License. Both models share the same 1,048,576 token context window.
Benchmarks
BenchGecko has no independent benchmark scores for Hy4 yet. Tencent's model card says a notable gap remains between Hy4 and top proprietary models on research level mathematics, long horizon coding and advanced visual question answering. Scores will appear on the model page once public leaderboards report them.
Source: BenchGecko · Tencent Hy4 model card on Hugging Face.
Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.