Hy3 Preview

The first Hy3 preview after Tencent rebuilt its training stack, combining fast/slow reasoning with sparse MoE.

LegacyReasoningCodingAgentOpen weights

MODEL NUMBER120

Specifications

Only published specifications are shown.

Total parameters

295B

Active parameters

21B

Architecture

MoE

Context

256K

Input modalities

Text

Output modalities

Text

Open weights

Yes

License

Apache-2.0

Status

Legacy

The first Hy3 preview after Tencent rebuilt its training stack, combining fast/slow reasoning with sparse MoE.

PreviousHunyuan Large389B · — ctx

CurrentHy3 Preview295B · 256K ctx

New capabilities Hybrid reasoning Tool use

Hybrid reasoningCodingTool use

2026-04-24Preview

Tencent released Hy3 Preview.

Select a card to view the related model.