A translation model that fits in 440 megabytes, runs entirely offline, and handles 33 languages just landed on Hugging Face, courtesy of Tencent. The model, called Hy-MT1.5-1.8B-1.25bit-GGUF, supports 1,056 translation directions, including five dialect and minority language variants, and requires no cloud connection, no API key, and no account. It is one of the most aggressively compressed machine translation models ever released publicly, packing 1.8 billion parameters into a file smaller than many smartphone games.
The release, which appeared on Tencent’s official Hugging Face organization page in early 2026, arrives at a moment when on-device AI is becoming a competitive battleground. Google already offers downloadable language packs for its Translate app, typically 40 to 50 megabytes per language pair. Apple’s built-in Translate supports offline use but covers only about 20 languages. Tencent’s approach is different: a single 440MB download that covers all 33 languages at once, with every possible pair available for translation.
How 1.8 billion parameters fit in 440 megabytes
The compression relies on two techniques developed under Tencent’s research umbrella. The first is a toolkit called AngelSlim, described in a February 2026 preprint as a framework for post-training quantization that pushes models into ultra-low-bit regimes. The second is a method called Sherry (arXiv:2601.07892), which achieves 1.25-bit quantization through a technique known as 3:4 fine-grained sparsity packing. In practical terms, Sherry fits four weight values into just five bits of storage, a ratio that would have seemed impractical for production-quality models even a year ago.
The result is a GGUF-format file, a format widely used in the llama.cpp ecosystem for running large language models on consumer hardware. GGUF files can be loaded on laptops, desktops, and smartphones without specialized inference servers, which makes the model immediately testable by anyone with a compatible device.
The training pipeline behind HY-MT1.5
The compressed model descends from Tencent’s HY-MT1.5 family, documented in the company’s Hugging Face model card and its GitHub repository for Hunyuan Translation. The family includes two sizes: a 1.8-billion-parameter variant (the one compressed here) and a larger 7-billion-parameter variant.
A technical report published on arXiv (ID 2512.24092) describes the training process in four stages: machine-translation-oriented pretraining on large parallel corpora, supervised fine-tuning on curated translation pairs, on-policy distillation to transfer knowledge from larger teacher models, and reinforcement learning to refine output quality. The report includes evaluations against Flores-200, WMT25, and Mandarin-to-minority-language benchmarks, though those scores apply to the full-precision model, not the 1.25-bit compressed version.
What developers still don’t know
For all its technical ambition, the release leaves several important questions unanswered.
Translation quality after compression. No published document from Tencent reports how much accuracy the model loses when squeezed from full precision to 1.25 bits per weight. The Flores-200 and WMT25 scores in the technical report were measured on the uncompressed model. Whether the 440MB variant introduces systematic errors, such as dropped negations, garbled named entities, or degraded performance on lower-resource languages, remains untested in any public evaluation.
On-device performance. Tencent has not published inference speed, memory consumption, or thermal behavior data for the model running on actual phones or tablets. A 440MB file fits easily on modern smartphones, but real-world latency depends on CPU architecture, available RAM, and how the inference runtime handles the unusual 1.25-bit weight format.
Licensing. Neither the Hugging Face model card nor the GitHub repository specifies clear commercial licensing terms. The AngelSlim project page, which hosts the base weights that Tencent’s model card references, also lacks an explicit license. Developers considering integration into shipping products will need to treat the weights as research-only until Tencent clarifies redistribution and monetization rights.
Corporate context. Tencent has not issued a press release, blog post, or executive statement accompanying the release. There is no public product roadmap indicating whether the model will be integrated into WeChat, QQ, or other Tencent products, or whether AngelSlim compression will be applied to other model families.
How to evaluate it yourself
The most direct way to assess the model is to download the GGUF file from Hugging Face, load it in a local inference framework like llama.cpp, and run translations on language pairs relevant to a specific use case. No API key or cloud account is needed.
Teams considering production deployment will likely need to build a custom evaluation harness. That typically means assembling a test set of sentences or documents in the relevant source languages, drawn from the actual product domain (news, e-commerce, customer support, technical documentation), and comparing the model’s output against both human reference translations and established systems like Google Translate or DeepL. Side-by-side human review is especially important for catching the kinds of subtle errors that automated metrics like BLEU often miss.
All three supporting papers (the HY-MT1.5 technical report, the AngelSlim toolkit paper, and the Sherry quantization paper) are hosted on arXiv, a preprint server supported by Cornell University. None have undergone formal peer review, a standard caveat worth noting when evaluating the technical claims.
Why this release matters beyond translation
Even with its open questions, the release is significant for what it signals about the trajectory of on-device AI. A year ago, running a 1.8-billion-parameter model on a phone would have required either heavy cloud offloading or a much larger local footprint. Compressing it to 440MB while maintaining (according to Tencent’s claims) usable translation quality across 33 languages suggests that the toolkit behind the compression, not just the translation model itself, could reshape how other AI capabilities are deployed on consumer hardware.
For researchers, the openly downloadable weights offer a concrete testbed for studying ultra-low-bit quantization in a real application rather than on synthetic benchmarks. For developers building multilingual products, particularly in regions with unreliable connectivity, a fully offline translation engine that covers this many languages in a single small package is worth serious evaluation, provided the quality holds up under testing.
The burden of proof now shifts to the community. Tencent has published the weights and the papers. Whether this 440MB model translates well enough to trust with real users is a question only independent benchmarking can answer.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
