Back to Gap Map Model components / Base / pretrained models

Hunyuan (Tencent Hy)

Tencent
open weights / Overall score: 4.0(strong)

Hunyuan is Tencent's large language model line, released under the Tencent Hy name since 2026. Hy3 is a 295B-parameter mixture-of-experts model with 21B active parameters and a 256K-token context, and Hy4 preview scales the design to 770B total and 49B active parameters with a one-million-token context. Both ship as instruct checkpoints with FP8 variants, aimed at coding, agent and office work.

Scored on the Hy releases. Tencent's video and 3D models are separate entries, hunyuan-video and hunyuan-3d.

Openness

3 high confidence
3.0
weights
open (Hy4 preview and Hy3, with FP8 variants, ungated on Hugging Face)
data
closed (no pretraining corpus or data listing is published)
code
partial (fine-tuning and reinforcement-learning guides only, no pretraining pipeline)
license
Apache-2.0 (OSI)
prior-release
distributed (Hunyuan-A13B, an earlier release, ships under Tencent's own Hunyuan-A13B license, not Apache-2.0)

Hy4 preview and Hy3 ship their weights under Apache-2.0, ungated on Hugging Face, with fine-tuning and reinforcement-learning recipes alongside. The pretraining data and the pretraining pipeline are not published, which keeps the line at open weights.

Adoption

4 high confidence
4.0

Hugging Face downloads summed across the Hy3 and Hy4 preview checkpoints and their FP8 variants, nearly all of them for Hy3.

Capability

4 low confidence
4.0

Tencent's own blind evaluations put Hy4 preview level with GLM 5.3 and Kimi K3 on engineering work, but its public benchmark tables are published only as images, and the vendor calls the release an early version with known issues. Without an independent result it sits a step below GLM rather than beside it.

  • https://huggingface.co/tencent/Hy3/raw/main/README.md recorded 2026-11-12

    Card: Hy3 is "a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters"; in a blind evaluation by 270 experts it reports Hy3 at 2.67 out of 4, "outperforming GLM-5.1 at 2.51/4". Benchmark charts are embedded as images.

  • https://huggingface.co/tencent/Hy4-preview/raw/main/README.md recorded 2026-11-12

    Card: Hy4 preview "comprises 770B total parameters, of which 49B are activated per token"; "163 internal experts rated model outputs on 203 engineering tasks. Hy4 preview came out slightly ahead of both GLM 5.3 (2.99 vs. 2.92 ...) and Kimi K3 (2.99 vs. 2.94 ...)"; and "This is an early version of Hy4 ... we are shipping with known issues".

Verified 2026-11-12