Descriptions:
Fahd Mirza tests Tencent’s Hy3 Preview, a 295 billion parameter mixture-of-experts model that activates only 21 billion parameters at inference time — making it far more efficient to serve than its raw size implies. The model ships with a 256k context window, an adjustable reasoning mode, and strong benchmark scores on coding and agentic tasks including competitive results on international math olympiad problems and Chinese PhD-level math qualifying exams.
Mirza runs three tests: a Langton’s Ant simulation built as a self-contained HTML file with canvas rendering, controls, and multi-ant support; a research agent task tracing nutmeg’s historical trade route from Banda Island through to Granada with internet search enabled; and a general capabilities sweep. The simulation output is visually clean with working sliders and counters. The historical research task produces a detailed, factually grounded paragraph covering the full spice trade chain without hallucinated shortcuts — a result Mirza specifically highlights as impressive.
Mirza is candid that previous Tencent models haven’t matched labs like Cohere or DeepSeek in his testing, but calls Hy3 Preview a meaningful step forward — particularly on instruction-following and long-context document tasks. For developers evaluating open-weight alternatives for production coding or research-agent pipelines, Hy3 Preview’s efficient MoE architecture and 256k context handling make it a model worth benchmarking directly against current Western open-source leaders.
📺 Source: Fahd Mirza · Published April 23, 2026
🏷️ Format: Review
