.png)
Tencent’s Hy Team released Hy3, a 295B-parameter Mixture-of-Experts model with 21B active parameters and a 256K context window. The launch follows a Hy3 Preview in late April, after post-training was scaled up using feedback from 50+ products.
Hy3 uses 192 experts with top-8 routing, which puts it in the class of sparse models that trade a much larger parameter count on paper for a smaller active footprint at inference time. Tencent says the model has 3.8B MTP layer parameters, and the public artifact size makes the deployment cost concrete: the full model is 598GB on Hugging Face, while the FP8 quantized version is 300GB.
That footprint matters more than the headline parameter count. For production teams, the relevant questions are whether the routing quality holds up under your traffic mix, how stable the model is under long-context workloads, and whether the quantized checkpoint can run efficiently enough to justify the extra orchestration complexity. On those dimensions, Hy3 is clearly aimed at serious hosted or multi-node self-hosted deployments rather than casual local experimentation.
Tencent claims Hy3 outperforms similar-size models and competes with flagship open-source models that have 2–5x more parameters. That’s the right comparison class for an MoE model: if the active parameter budget is strong, it can land close to much larger dense models in quality while avoiding a proportional inference cost increase.
The stronger claim is the “significant gains in utility across various products and productivity tasks” after gathering feedback from 50+ products during the preview period. That suggests Tencent optimized for practical task behavior, not just benchmark chasing. I’d still treat any broad utility claim as a prompt to test on your own workload, especially if your evaluation includes long-document editing, multi-step tool use, or brittle formatting tasks where model preferences can differ sharply from standard text benchmarks.
For teams choosing between Hy3 and a large dense open model, I’d benchmark three things first: long-context stability at 128K+ tokens, structured output reliability, and latency under expert-routing load. If Hy3 is materially better on those, the extra operational complexity becomes easier to justify.
Hy3 is available for free on OpenRouter until July 21st, which makes it easy to evaluate without arranging direct model access. That matters because the practical bar for adoption is no longer “is the model good in isolation,” but “is it good enough to displace an existing production model for a specific slice of traffic.”
The quickest way to judge it is to run your own internal eval set against the tasks that hurt you today: long-context retrieval, agentic tool calls, code transformation, and any domain-specific generation where formatting and instruction adherence matter. Tencent is clearly targeting those use cases, but the only meaningful test is whether Hy3 reduces escalation rates or improves success rates in your own pipeline.
The early preview already drew attention for ranking strongly on OpenRouter, and the current release appears to be a more mature version of that same model family. That makes Hy3 worth a serious look if you need a high-capacity long-context model and can absorb the infrastructure cost of a very large MoE checkpoint.
