Tencent open-sources HYMT2 translation models that run on phones

HYMT2 is a family of open-source multilingual translation models from Tencent, spanning from 1.8B and 7B up to 30B parameters. Designed to support 33 languages, its standout capability is strict instruction following, ensuring structural tags, formatting, and code placeholders remain completely intact during translation.

Built for complex production formatting

Translating real-world enterprise content is rarely about plain sentences. In practice, localization involves product web pages, subtitles, app UI strings, markdown files, and deeply nested JSON payloads. A single broken delimiter or translated variable name can break an entire build.

HYMT2 is fine-tuned specifically to handle these constraints. Beyond simply changing languages, it can follow detailed system prompts to preserve custom formatting, stick to rigid domain glossaries, keep delimiters intact, or selectively translate only user-facing strings within structured schemas.

In instruction-following benchmarks, the HYMT2 series outperformed significantly larger models, including DeepSeek V4. Across domain-specific evaluations covering finance, legal, medical, tech, and politics, HYMT2 similarly beat open models that are dozens of times larger.

Hardware specs and deployment

The flagship 30B model uses a Mixture-of-Experts (MoE) architecture. While the overall footprint sits around 60GB, it activates just 3 billion parameters per token during inference, keeping throughput high.

For local and edge setups, Tencent also released a 7B variant and an ultra-compact 1.8B version. The 1.8B weights take up around 4GB of storage, fitting comfortably on consumer GPUs and modern mobile phones. Both FP8 and GGUF quantized releases are available out of the box, with setup instructions provided for local machine deployment.

Why it matters

If your development workflow handles internationalization files, subtitles, or programmatic JSON translation, HYMT2 is an immediate candidate for local testing. It solves a notoriously frustrating pipeline issue without burning API credits on giant frontier models.

Tencent's strategy here is sharp and practical: instead of fighting for generic leaderboard spots, they packaged professional, structure-preserving translation into an efficient edge package. It gives developers a reliable offline tool that does exactly what the system prompt asks.