Summary
Sam Witteveen breaks down Tencent’s newly-released Hy3, a 295-billion-parameter mixture-of-experts model with 21 billion active parameters and a 3.8-billion-parameter speculative decoding companion model. After months in preview, the full release marks what Witteveen describes as Tencent formally declaring its intent to compete as a frontier lab — and to do so in the open-source ecosystem, a departure from the company’s historically internal-focused AI development.
The video is candid about where Hy3 sits competitively: it does not beat GLM 5.2 on most agentic coding benchmarks, and Qwen 3.7 Max holds advantages in several reasoning tasks. Where Hy3 does shine is efficiency — at roughly half the parameter count of GLM 5.2 while posting competitive scores against DeepSeek-class models, making it a credible option for enterprises wanting a single fully-local model that can run on obtainable hardware without requiring racks of B200s. The 256K context window is notably smaller than GLM 5.2’s, which Witteveen flags as a real limitation.
Live testing on OpenRouter (currently free for the first two weeks post-launch) covers SVG generation, long-form writing, and structured output tasks, including a side-by-side comparison with the earlier preview version showing measurable training improvements. The verdict: Hy3 is not the right choice for pure coding workloads, but for companies looking for a capable, locally-deployable, agentic-task-oriented model in the 200B+ class, it is now a serious contender worth evaluating.
📺 Source: Sam Witteveen · Published July 07, 2026
🏷️ Format: Review
