Tencent open-sources its flagship 770B-parameter MoE model Hy4-preview

🤖 Tencent open-sources its 770B-parameter flagship Hy4-preview

The Tencent Hy / Hunyuan team has open-sourced the MoE model Hy4 preview: 770B parameters, 49B active per token, context up to 1M tokens, Apache 2.0. The claimed GPQA Diamond score is 92.3. BF16 and FP8 weights are available on Hugging Face, ModelScope, GitCode, and CNB.

🌍 The second open frontier flagship under Apache 2.0 in a short period: IndexCache reduces inference costs on contexts up to 1M tokens, an MTP layer enables speculative decoding, and ready-made vLLM and SGLang images allow you to deploy the model yourself.

👤 The weights can be run on 8 GPUs via vLLM or SGLang through an OpenAI-compatible API. Reasoning is enabled by default and can be disabled with reasoning_effort=no_think. The early version spends extra time on re-verification — faster iterations are promised.

Source 1: https://www.tencent.com/tencent-releases-and-open-sources-tencent-hy4-preview/

Source 2: https://huggingface.co/tencent/Hy4-preview

Sources