Hy-MT2-1.8B

According to the ThursdAI Tencent company page, Tencent's Hunyuan lab open-sourced the Hy-MT2 translation model family under Apache 2.0 on May 28, 2026, with two variants: a tiny 1.8B model and a larger 30B-A3B MoE sibling. The page attributes the 1.8B variant as reportedly beating paid translation APIs such as Microso The same ThursdAI page contextualizes Hy-MT2 within Tencent's broader Hy family output, noting 18 Tencent releases since January 2025 across 3D, video, avatar, OCR and portrait models, with 13 shipping open weights. Coverage is presented as a podcast/episode recap aggregated by ThursdAI rather than as an official Tence

BenchmarkList provides model-specific evaluation data for Hy-MT2-1.8B, listing self-reported scores dated May 26, 2026 across five benchmarks. These include MAXIFE at 56.17% (general multilingual instruction following), IFEval at 80.22% (instruction following with verifiable constraints), IFBench at 35.3% (complex inst The page lists a pricing row of $0.044 per 1M input tokens and $0.177 per 1M output tokens for Hy-MT2-1.8B. The IFMTBench entry describes its methodology: 4,506 single-constraint and 2,838 multi-constraint items across six constraint dimensions and five compositional patterns, with instructions issued in all seven lang

Tencent Hunyuan open-sourced its Hy-MT2 multilingual translation model family on May 21, 2026, with the Hy-MT2-1.8B variant explicitly named as a lightweight mobile-optimized model. The model uses Tencent's self-developed AngelSlim 1.25-bit extreme quantization to compress to roughly 440MB, enabling local offline infer The Hy-MT2 family covers bidirectional translation across 33 languages plus five ethnic languages/dialects, with sibling variants Hy-MT2-7B (balanced speed/performance) and Hy-MT2-30B-A3B (a new mixture-of-experts flagship). According to the report's cited official evaluations, Hy-MT2-1.8B reached 88.1% of Gemini 3.1 P