Tencent’s Hunyuan team has open-sourced extreme quantized builds of its Hy-MT1.5-1.8B translation model aimed at phones: Hy-MT1.5-1.8B-1.25bit at about 440 MB and Hy-MT1.5-1.8B-2bit at about 574 MB. Both landed on Hugging Face on April 29, 2026, with weights and GGUF-oriented packaging for local use.

The pitch is straightforward. Full-precision Hy-MT1.5-1.8B needs roughly 3.3 GB in FP16—awkward on mid-range handsets. Cutting parameter storage from 16-bit down to 2-bit or 1.25-bit is meant to keep translation quality while fitting offline, private inference on ordinary mobile CPUs.

What's new

Hy-MT1.5-1.8B is a dedicated translation model, not a general chatbot with a translate prompt glued on. Tencent says it was trained through an MT-oriented pipeline that includes pre-training, supervised fine-tuning, on-policy distillation, and reinforcement learning. Official materials put coverage at 33 languages, five dialects / minority languages, and 1,056 translation directions.

Languages and dialects listed across the Hunyuan-MT / Hy-MT family cards include Chinese, English, French, Portuguese, Spanish, Japanese, Turkish, Russian, Arabic, Korean, Thai, Italian, German, Vietnamese, Malay, Indonesian, Filipino, Hindi, Traditional Chinese, Polish, Czech, Dutch, Khmer, Burmese, Persian, Gujarati, Urdu, Telugu, Marathi, Hebrew, Bengali, Tamil, Ukrainian, Tibetan, Kazakh, Mongolian, Uyghur, and Cantonese.

The new cards are quantizations of that 1.8B specialist:

  • Hy-MT1.5-1.8B-1.25bit — ~440 MB; Sherry ternary scheme; STQ-oriented mobile path
  • Hy-MT1.5-1.8B-2bit — ~574 MB; heavier bit-width for devices that can spare the RAM

Sherry, accepted at ACL 2026, is the 1.25-bit method: for every four weights, the three most important are stored as 1-bit values ({−1, +1}) and one is zeroed, packing four weights into five bits. Tencent pairs that with AngelSlim compression tooling and a custom STQ kernel aimed at mobile SIMD alignment; a related llama.cpp path was also floated for community builds.

Company evaluations claim the lightweight 1.8B line can match or beat mainstream commercial systems such as Google Translate on cited benchmarks—including Flores-200 Chinese–foreign mutual translation charts on the model cards—and in some dimensions compete with much larger general models. Those comparisons are Tencent’s own; independent third-party phone A/B results are not part of this release package.

A ready Android demo highlights background word selection—translate from mail, local pages, or chat without switching apps—with inference kept on-device so text need not leave the phone. Tencent also says the capability is already used inside some of its own meeting, office, and browser surfaces.

Full-precision Hy-MT1.5-1.8B loads via Transformers on the model card; the 1.25-bit phone build follows the STQ / llama.cpp path documented there.

Where the larger Hunyuan-MT models fit

This is not “the WMT champion 7B squeezed into a phone.” Hunyuan-MT-7B and Hunyuan-MT-Chimera-7B are an earlier open release (September 2025). The 7B model is the main translator; Chimera is an ensemble “slow thinking” fusion model that combines multiple 7B candidates into one refined output. Together they led the WMT2025 general MT shared task in 30 of 31 language pairs they entered.

Hy-MT1.5-1.8B sits later in the same product family as a smaller specialist. The April 2026 news is the on-device quantization of that 1.8B line—not a direct shrink of Chimera or the 7B checkpoint into 440 MB.

Why it matters

Most strong multilingual MT still assumes a cloud API. A sub-half-gigabyte pack that claims competitive quality changes the default for travel, restricted networks, and privacy-sensitive reading: download once, translate forever, keep the plaintext local.

It also sharpens the open-weight race around deployable size, not just leaderboard FLOPs. If 1.25-bit MT holds up outside vendor charts, “good enough offline translate” stops being a niche engineering demo and becomes a packaging target other labs will have to match.

Our take

The credible story is productization: a trained 1.8B MT model plus a hardware-aware 1.25-bit stack, not a mythic 7B→pocket miracle. Treat the vendor Flores-200 and commercial-API wins as a starting claim until independent on-device latency and quality numbers land. The interesting bet is whether Sherry-class quantization becomes the default shipping format for niche models that already “won” at full precision.

Sources