Hy-MT2-30B-A3B

Tencent released the Hy-MT2 family of translation models under Apache 2.0, including a 30B-A3B MoE variant, as covered on ThursdAI's multilingual topic page dated May 28, 2026. The family also includes a 1.8B sibling reported to outperform paid translation APIs such as Microsoft Translator, which the podcast framed as The Hy-MT2 release positions Tencent as shipping competitive translation-capable open weights, with the 30B-A3B MoE variant serving as the larger-capacity option alongside the compact 1.8B model. The model subject here is Tencent's Hy-MT2-30B-A3B itself, not any gateway's resale of it, and the excerpt attributes creati

The Hy-MT2 paper on alphaXiv describes a family of fast-thinking multilingual translation models from Tencent's Hunyuan team, with three sizes — 1.8B, 7B, and 30B-A3B MoE — all supporting translation across 33 languages and following translation instructions in multiple languages. The authors position the family as han The paper reports that the 7B and 30B variants outperform open-source models such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode, while the lightweight 1.8B surpasses mainstream commercial translation APIs from Microsoft and Doubao overall. For on-device deployment, AngelSlim's 1.25-bit extreme quantization com

The NVIDIA NeMo AutoModel 0.5.0 release notes (dated 26.06) add native model support and a supervised fine-tuning recipe for HY-MT2-30B-A3B. The same release broadens coverage to other large MoE and dense models — DeepSeek-V4 Flash, ERNIE 4.5, MiMo-V2-Flash, Ling 2.0, Hy3-preview, Falcon H1, MiniCPM5-1B, and Nemotron-3 For Hy-MT2-30B-A3B specifically, NeMo AutoModel 0.5.0 ships a native model and checkpoint adapter plus an SFT recipe, enabling teams to fine-tune the 30B-A3B translation MoE within NeMo's training pipelines. The release also extends discrete-diffusion LLM training to LLaDA2 and Nemotron-Labs-Diffusion, adds diffusion f

ThursdAI's Tencent releases index tracks 19 Hunyuan-team releases since January 2025, 14 shipped with open weights, including the Hy-MT2 translation family on May 28, 2026 under Apache 2.0. The same Hunyuan portfolio covers 3D generation, video, avatars, OCR, and portrait models, with more recent entries such as the AU Earlier 2026 entries include the Hy3-preview 295B/21B MoE reasoning model and the Hunyuan3D WorldClaw text-to-3D game-world paper (Aug 13, 2026). The index positions Hy-MT2 within Tencent's broader pattern of releasing specialized, open-weight models for distinct modalities — translation, speech, 3D, reasoning — with H

Tencent Hunyuan has open-sourced Hy-MT2, a multilingual translation model family, and launched a companion "Tencent Hy Translation" mini program based on it. The family ships in three sizes — a lightweight 1.8B variant, a 7B "sweet spot" model, and the flagship Hy-MT2-30B-A3B — all supporting bidirectional translation Hy-MT2-30B-A3B is the family's first mixture-of-experts design, expanding total parameter count while keeping activated parameters per inference in check for professional-scenario translation. Tencent's published evaluations report the 1.8B/7B/30B-A3B models reaching 88.1%, 96.9%, and 98.1% of Gemini 3.1 Pro's level on

BenchmarkList's dedicated Hy-MT2-30B-A3B page reports self-reported evaluation scores dated 2026-05-26 that align with the May 28, 2026 release. On the multilingual MAXIFE benchmark the model scores 77.4% (33rd percentile, loose score 80.5%, strict 74.3%); on IFEval it reports a final score of 89.8% at 79th percentile, The page also documents a new IFMTBench evaluation designed for multilingual translation instruction following, covering seven languages with thousands of single- and multi-constraint items across six constraint dimensions. Pricing is listed at $0.074 per 1M input tokens and $0.295 per 1M output tokens. All scores are