Hunyuan 3.0 (Hy3)

Positioning

Don't let 21B active parameters fool you: MoE routing needs every expert resident in memory, so this is a datacenter model. BF16 needs roughly 590GB of VRAM and the official FP8 checkpoint is about 300GB before KV cache (which adds 80–120GB at long context) — Tencent's own recommendation is an 8x H20-3e node. Even a 512GB Mac Studio would burn over half its unified memory on the FP8 weights alone. There is no sanctioned single-box story here.

Who should run it: teams with 8x H100/H200-class hardware that want an Apache 2.0 agent backbone. The license is the real headline — the April preview's territorial restrictions (including the EU/UK exclusion) are gone, with no registration or acceptable-use strings, which puts it ahead of Llama for any org where legal review is the bottleneck. Reliability engineering is the differentiator: sub-4% tool-call variance across scaffoldings, hallucination rate cut from 12.5% to 5.4% in Tencent's internal evals, and 47.4% fewer tokens than GLM-5.2 on agent tasks — which compounds into real serving savings.

Who should skip it: coding-first users — GLM-5.2 (753B / ~40B active, double the compute per token) still leads there — and anyone doing frontier math (MathArena Apex 38.7 vs GPT-5.5's 85.4). And obviously anyone without a rack. For agent workloads at scale, this is arguably the cleanest-licensed deal in open weights right now.