Hy3 preview matters because modern agent and reasoning tasks increasingly demand models that can (a) hold extremely long contexts and (b) scale selective compute via experts. This release packages a MoE model trained on a rebuilt Tencent infrastructure and made publicly available to let practitioners test MoE trade-offs for instruction, coding, and agent-style workloads.
Key Capabilities
- Reasoning & STEM: reported strong results on hard benchmarks (math/coding/STEM) indicating improved chain-of-thought and multi-step problem solving compared with many base open models — useful when accuracy on multi-step tasks matters.
- Long-context & instruction following: native support for very long contexts (document-level workflows) and an instruct-tuned variant for multi-turn conversational and instruction-following tasks, which helps when tasks require carrying large context windows or complex rule sets.
- Code & agent workflows: explicit engineering for coding and agent benchmarks; pretrained weights and recommended inference recipes (vLLM/SGLang) target integrations where models must interact with tools, terminals, or multi-stage agents.
- Deployment & compatibility: model artifacts are provided in safetensors and Transformers-compatible formats on Hugging Face; recommended inference backends are mentioned in the model card to enable practical serving experiments.
Who it's for and trade-offs
Great fit if you: want to experiment with MoE at scale for long-context reasoning or agent pipelines; need an instruct-capable LLM with demonstrated coding and STEM strengths; or are evaluating trade-offs between activated-parameter efficiency and total-parameter scale.
Look elsewhere if you: need a drop-in, low-cost hosted endpoint (Hy3-preview requires careful infra and multi-GPU/MTP support), require permissive open-source licensing for commercial redistribution (Hy3 preview uses the Tencent Hy Community License), or must run on single modest GPUs — MoE models impose extra memory/serving complexity and routing requirements.
Where it fits
Use this model when comparing MoE routing strategies, evaluating long-context instruction-following at scale, or building agent systems that benefit from stronger multi-step reasoning. For lightweight or latency-sensitive production deployments, prefer smaller dense models or distilled variants until serving cost and routing are solved for your infra.
(Release note: model weights were published to Hugging Face on 2026-04-13 and open-sourced by the Tencent Hy Team; consult the model card for benchmark tables, quantization notes, and the Tencent Hy Community License.)
