ZMIME

Hy4 Base vs TML-d24-varloop-research

Compare documented capabilities and access. A dash means a comparable figure is unavailable.

Model specifications
SpecificationHy4 BaseTencentTML-d24-varloop-researchThinking Machines
TaskMultimodal pretrainingVariable-depth reasoning research
Published capabilitiesPretrained checkpoint for continued training and adaptation. Tencent’s flagship Hy4 uses a 78-layer language backbone with a separate vision encoder.Researchers can vary the number of middle-layer passes with a single checkpoint.
Model size770B total / 49B active per token in the language backbone26.4B total / 3.4B active parameters
AccessDownloadable under the Tencent Hy LicenseDownloadable research checkpoint
Operating limits—Native mixed-precision checkpoint occupies approximately 20 GB.
SourcesModel weightsModel weights · Additional documentation
InputsText, Images, VideoText, Images
WeightsOpen weights · Tencent Hy LicenseOpen weights · Apache-2.0
ReleasedNov 20, 2026Nov 20, 2026