Tencent Releases Hy4-Preview 770B Model Weights with 49B Active Parameters

Tencent released open weights for Hy4-preview, a 770B-parameter MoE model with 49B active parameters, on HuggingFace. This is the first frontier-scale open-weight release from a major Chinese lab at this size tier.

For operators, this shifts the baseline for self-hosted capability. If you can provision multi-node inference with ~300GB+ VRAM for active weights plus expert shards, you now have a credible alternative to Western API dependencies for reasoning-heavy workloads. The practical cost is electricity and engineering time, not per-token licensing. Expect pressure on closed-API pricing for equivalent intelligence tiers.

Operationally, this makes two things possible: first, fine-tuning or continued pretraining on a genuinely large MoE without API constraints — relevant for domain adaptation under data sovereignty rules. Second, it normalizes serving 49B-active MoEs on commodity 8xH100 or similar clusters, which was previously reserved for a few vendors. The immediate workflow change is that evaluation suites must now include this model in offline baselines, especially for Chinese-language tasks and math/coding benchmarks where Tencent has optimized. If you lack hosting capacity, the strategic question is whether to wait for quantized or distilled variants rather than ignore the release.