On 21 May 2026, Tencent's Hunyuan team released Hy-MT2, a new generation of multilingual translation models aimed at production workloads—not another re-packaging of the April 2026 Hy-MT1.5 phone quantizations. The family ships in three sizes: a dense 1.8B, a dense 7B, and a 30B-A3B mixture-of-experts (MoE) variant with about three billion active parameters per forward pass.
All three checkpoints target the same broad coverage as earlier Hy-MT releases—33 languages and 1,056 translation directions—but Hy-MT2's headline upgrade is instruction following: terminology locks, style constraints, delimiter preservation, and structured formats (JSON, YAML, code snippets) where keys, placeholders, and markup must stay untranslated.
Related: For the earlier story on extreme 1.25-bit / 2-bit Hy-MT1.5-1.8B phone builds (~440 MB offline), see our April coverage of the ultra-light Hy-MT1.5 packs. Hy-MT2's 1.8B line inherits that on-device path; this article focuses on what is new in the Hy-MT2 generation—especially 7B, MoE, and the instruction layer.
What's new
- Three Hy-MT2 sizes: 1.8B dense, 7B dense, and 30B-A3B MoE—each released as open weights on Hugging Face / ModelScope with GGUF builds surfacing via community channels such as Ollama.
- Instruction-following templates: Built-in patterns for default translation, glossary / terminology intervention, style control, personalization, delimiter retention, and structured-data translation with strict non-translation rules for keys, tags, and placeholders.
- IFMTBench: A new benchmark Tencent published alongside Hy-MT2 to score instruction-constrained translation—not just plain sentence-pair quality.
- Benchmark claims: Tencent reports the 7B and 30B-A3B models ahead of open rivals such as DeepSeek-V4-Pro and Kimi K2.6 in fast-thinking mode on selected suites, plus WMT25, FLORES-200, domain-specific (finance, law, medicine), and business-style corpora.
- Integration assets: Hy-MT2-Translator Skill and a dedicated
Tencent-Hunyuan/Hy-MT2repository separate from the older Hy-MT1.5 tree. - WMT26 partnership: Tencent is co-stress-testing the family on video-subtitle and general MT tasks—a signal the line is meant for competitive, visible deployment—not lab charts only.
How this differs from Hy-MT1.5
Hy-MT1.5 (covered in our prior piece) was primarily about making a specialist 1.8B translator deployable on phones via Sherry / AngelSlim quantization. Hy-MT2 keeps a 1.8B entry point but reframes the product around controllable, template-driven translation and adds 7B and 30B-A3B MoE options for teams that need higher ceiling quality or MoE efficiency—not just pocket-sized inference.
If your reader cares only about offline 440 MB phone inference and Sherry mechanics, the 1.5 article remains the right entry. If they care about production MT with glossary/style/JSON rules or MoE serving economics, Hy-MT2 is the relevant release.
Why it matters
Most cloud MT APIs treat every request as plain text in, plain text out. Hy-MT2's instruction layer is aimed at pipelines where legal, medical, localization, and developer workflows need deterministic constraints—fixed terms, brand voice, or file formats—without brittle post-processing regex stacks.
The MoE variant is the other strategic bet: near-large-model quality with fewer active parameters per token, if serving memory and routing overhead stay manageable in real deployments. The 7B dense checkpoint is the pragmatic middle tier for teams that outgrew 1.8B but do not want MoE ops complexity.
Our take
The credible Hy-MT2 story is product generation + control surface, not another "tiny model beats Google" phone headline—that ground was already covered by Hy-MT1.5. Watch whether IFMTBench scores translate into reliable behavior on messy real inputs (mixed markup, partial JSON, inconsistent glossaries) and whether the 30B-A3B MoE's latency and RAM profile are acceptable outside Tencent's own charts. If both hold, Hy-MT2 is a meaningful step from "good offline translate" to "translation you can actually pipe into software workflows."
