# Hy-MT2-1.8B ![Organization](https://img.shields.io/badge/Org-Hungyuan-6366f1) ![Context](https://img.shields.io/badge/Context-262K%20tokens-1f6feb) ![License](https://img.shields.io/badge/License-Apache%202.0-2da44e) ![Modality](https://img.shields.io/badge/Modality-Text-8957e5) ![Released](https://img.shields.io/badge/Released-May%202026-555) > **Quick answer:** Hy-MT2-1.8B is the lightweight member of Tencent Hunyuan's Hy-MT2 family — a 1.8B-parameter, Apache 2.0-licensed multilingual translation-specialist model supporting 33 languages. It scores 79.77 (XCOMET-XXL) on FLORES-200's full translation matrix and 50.30 on WMT25, and — despite its small size — outperforms commercial translation APIs from Microsoft and Doubao overall. With AngelSlim's 1.25-bit quantization it needs only ~440MB of storage, making it well suited for on-device deployment. ## At a Glance **Where Hy-MT2-1.8B leads** - Outperforms Tower-Plus-72B (a much larger open translation-specialist model) and is competitive with Microsoft Translator and Doubao Translator across FLORES-200, WMT25, and domain/real-world translation benchmarks - Extreme on-device efficiency: AngelSlim 1.25-bit quantization shrinks storage to ~440MB with a 1.5x inference speedup over the prior generation's 4-bit build - Supports 33 languages including several minority/low-resource Mandarin-adjacent language pairs **Where it lags** - Clearly behind its own larger siblings (Hy-MT2-7B, Hy-MT2-30B-A3B) on every reported benchmark, and behind frontier general-purpose models like Gemini 3.1 Pro on raw translation quality - General instruction-following (IFBench, IFEval, MaXIFE, Multi-IF) is notably weaker than translation-specific instruction following — it's tuned for translation tasks, not general-purpose chat **Best for:** On-device or resource-constrained translation deployments (mobile apps, embedded systems, edge inference) where a dedicated small translation model outperforms general-purpose LLMs of similar size. ## What Hy-MT2-1.8B Is Hy-MT2-1.8B is the smallest of three sizes (1.8B, 7B, 30B-A3B) in Tencent Hunyuan's second-generation "fast-thinking" multilingual translation model family, described in the technical report "Hy-MT2: A Family of Fast, Efficient and Powerful Multilingual Translation Models in the Wild." It's a dense transformer (32 layers, 2048 hidden size) trained via a three-stage Family-Centric Post-training (FCPT) pipeline: MT-oriented mid-training on ~1T tokens of translation-related data, followed by Reference-Guided On-Policy Distillation and Family-specific RL training organized around language families (Western European, East Asian, Middle Eastern right-to-left), then a final Cross-family distillation step into a unified model. Unlike general-purpose LLMs that treat translation as one of many capabilities, Hy-MT2-1.8B is purpose-built and evaluated across four dimensions: general translation (FLORES-200, WMT25, Mandarin⇔Minority), real-world business scenarios (Tencent's own WildMTBench), domain-specific translation (Tencent's own DomainMTBench, covering finance/law/medicine/technology/politics/education), and translation instruction-following (IFMTBench, plus general instruction benchmarks IFBench/IFEval/MaXIFE/Multi-IF). It also supports structured translation-instruction prompting — preserving delimiters, following style constraints, translating only specific fields in structured data, and respecting terminology references. ## Specifications | Field | Value | |---|---| | Organization | Hungyuan (Tencent) | | Architecture | Dense transformer | | Parameters | 1.8B | | Layers | 32 | | Hidden size | 2,048 | | Context length | 262,144 tokens | | Supported languages | 33 | | License | Apache 2.0 | | Release date | May 2026 | | Modality | Text | | Quantized variants | FP8, Q4_K_M (GGUF), 2-bit, 1.25-bit (AngelSlim, ~440MB) | ## Pricing Open weights under Apache 2.0 — self-hosted via transformers, vLLM, SGLang, or llama.cpp (GGUF builds available, including 2-bit and 1.25-bit AngelSlim-quantized variants for extreme on-device compression). No API list price; cost is your own inference infrastructure. ## Public Benchmark Scores | Benchmark | Score | Source | Date | |---|---|---|---| | [FLORES-200](/benchmarks/meta-ai/flores-200) | 79.77 (XCOMET-XXL) | Hy-MT2 technical report | 2026-05 | | [WMT25](/benchmarks/wmt/wmt25) | 50.30 (XCOMET-XXL) | Hy-MT2 technical report | 2026-05 | *Hy-MT2-1.8B also reports scores on Tencent's own in-house evaluation suites (WildMTBench, DomainMTBench, IFMTBench, Mandarin⇔Minority testset) and on general instruction-following benchmarks (IFBench, IFEval, MaXIFE, Multi-IF) — not added as Benchgen leaderboard entries since these either lack an independent public benchmark page (WildMTBench/DomainMTBench/IFMTBench are Tencent-constructed, non-independently-reproducible test sets) or measure general instruction-following rather than translation capability specifically.* ## Hy-MT2-1.8B vs Alternatives | Model | FLORES-200 (XCOMET-XXL) | License | Params | |---|---|---|---| | **Hy-MT2-1.8B** | 79.77 | Apache 2.0 | 1.8B | | [Hy-MT2-7B](/models/hungyuan/hy-mt2-7b) | 86.89 | Apache 2.0 | 7B | | [Hy-MT2-30B-A3B](/models/hungyuan/hy-mt2-30b-a3b) | 87.47 | Apache 2.0 | 30B (3B active) | Hy-MT2-1.8B trades peak translation quality for deployability — it's meaningfully behind both larger siblings on FLORES-200, but remains the only model in this family (and one of very few models at this parameter scale generally) purpose-built and independently competitive against commercial translation APIs. ## Related Models & Benchmarks - Compare: [Hy-MT2-7B](/models/hungyuan/hy-mt2-7b) · [Hy-MT2-30B-A3B](/models/hungyuan/hy-mt2-30b-a3b) - Benchmarks: [FLORES-200](/benchmarks/meta-ai/flores-200) · [WMT25](/benchmarks/wmt/wmt25) ## Frequently Asked Questions

What is Hy-MT2-1.8B?

Hy-MT2-1.8B is Tencent Hunyuan's smallest translation-specialist model, a 1.8B-parameter dense transformer supporting 33 languages, released under Apache 2.0.

Is Hy-MT2-1.8B open source?

Yes — it's released under the Apache 2.0 license with weights on Hugging Face and ModelScope, including FP8, GGUF, 2-bit, and 1.25-bit quantized variants.

How does Hy-MT2-1.8B compare to Hy-MT2-7B and Hy-MT2-30B-A3B?

Hy-MT2-1.8B trades translation quality for size — it scores 79.77 on FLORES-200 (XCOMET-XXL) versus 86.89 for Hy-MT2-7B and 87.47 for Hy-MT2-30B-A3B, but is far more efficient to deploy, especially with 1.25-bit on-device quantization (~440MB).

--- *Specs from Tencent Hunyuan's official Hy-MT2-1.8B Hugging Face model card and technical report (arXiv:2605.22064). Last updated 2026-08-31.*

Hy-MT2-1.8B — Training jobs

Training jobs