Tencent drops Hy3 preview weights

Hy3 preview is a serious open-weights entrant, but its practical audience is narrower than the headline suggests: big-GPU operators, benchmark watchers, and teams comfortable with Tencent's custom license.

• –21B active parameters makes the MoE design efficient on paper, but serving still calls for 8 large-memory GPUs, so this is not a casual local model
• –Tencent is leaning into coding and agent evals, including SWE-bench Verified, Terminal-Bench 2.0, BrowseComp, and internal workflow benchmarks
• –The 256K context window and OpenAI-compatible serving path make it more interesting for long-context agent stacks than simple chat replacement
• –Community reaction is mixed: users like another open-weights option, but early comments flag restrictive licensing and coding performance that may trail smaller Qwen-class models
• –The release matters less as a single model drop than as evidence Tencent is rebuilding Hunyuan around product feedback loops and agent workloads