Tencent Hunyuan Open-Sources Lightweight AI Translation Model: Edge-Cloud Synergy Breaks Through Computing Power Bottlenecks

Amid the accelerating deployment of large AI models across industries, Tencent’s Hunyuan Lab officially open-sourced version 1.5 of its translation large language model on July 3, offering two parameter scales: 1.8B and 7B. This release marks a pivotal breakthrough in domestic AI infrastructure for lightweight deployment, with the 1.8B model leveraging innovative distillation techniques to achieve performance on mobile devices that approaches cloud-based large models.

The most striking aspect of this open-source release is the revolutionary improvement in computing efficiency. The 1.8B parameter model, after quantization and compression, requires only 1GB of memory and achieves an inference speed of 0.18 seconds per 50 tokens on Arm-based mobile chips — a 55% acceleration over the industry average. Notably, this performance is achieved without sacrificing accuracy: according to FLORES-200 benchmark tests, its translation quality reaches 90% of the level of hundred-billion-parameter models like Gemini 3.0 Pro, far surpassing other open-source competitors of similar scale. At the technical level, the pioneering “large model guiding small model” mechanism is a key breakthrough. Tencent employs On-Policy Distillation technology, where the 7B model provides real-time guidance to the 1.8B model in learning prediction deviation corrections — a departure from the static knowledge transfer used in traditional distillation. This dynamic learning mechanism enables a qualitative leap in the small model’s logical reasoning capabilities, offering a new paradigm for edge computing scenarios.

In terms of AI infrastructure compatibility, the model has been adapted to multiple chip architectures including Qualcomm, Intel, and Muxi, covering the full spectrum of computing needs from cloud GPUs to edge-device NPUs. Targeting vertical sectors such as healthcare and law, the newly added terminology customization feature solves the challenge of semantic consistency in specialized domains — signaling a shift from general-purpose AI translation toward industry-specific services. Language coverage exhibits clear “infrastructure” characteristics: it supports 33 major global languages while strengthening processing capabilities for low-resource languages like Czech. Support for five Chinese ethnic minority languages and dialects reflects a new trend in localized AI development. Combined with upgrades such as long-text context understanding, the model is redefining the standards of human-computer interaction in translation experiences.

总结