# HY-MT1.5 Technical Report Source: https://www.youtube.com/watch?v=QP36qdYvr1w Recap page: https://rapidrecap.app/video/QP36qdYvr1w Generated: 2026-01-19T17:05:08.941+00:00 --- ## Quick Overview The technical report for HY-MT1.5 demonstrates that the smaller, locally-run model significantly outperforms the larger, cloud-based Gemini 3.0 Pro model in translation tasks involving minority languages like Tibetan and Uger, achieving 99% of the quality at a fraction of the cost and latency, primarily by utilizing a specialized training pipeline focused on context preservation and avoiding common pitfalls like relying solely on simple metrics. **Key Points:** - The HY-MT1.5 model, despite being small enough to run locally on a phone, significantly outperformed Gemini 3.0 Pro in translation quality for minority languages like Tibetan and Uger. - The 1.8B parameter HY-MT1.5 model scored 2.74 on Chinese-English tasks, beating the 72B parameter Gemini 3.0 Pro model's score of 2.55 on the same benchmark. - HY-MT1.5 achieves this superior performance by using a five-stage training pipeline that includes specialized instruction tuning and incorporating reverse KL divergence to match the teacher model's probability distribution. - The paper explicitly notes that standard AI training relies too heavily on simple metrics, whereas HY-MT1.5 focuses on preserving context, nuance, cultural appropriateness, and avoiding the 'causal blunder' of misinterpreting the reason for an output. - The small model's ability to run locally allows for instant, private, and secure translation without high-speed internet, addressing the digital divide. - The training pipeline involves stages of pre-training, instruction tuning, reverse KL divergence optimization, and finally, reinforcement learning (RL) based on human feedback (RLHF) using a specialized rubric. ![Screenshot at 03:38: The speaker discusses how Reverse KL Divergence forces the student model to match the teacher's probability distribution, which is a key methodological difference in the training pipeline.](https://ss.rapidrecap.app/screens/QP36qdYvr1w/00-03-38.jpg) **Context:** This video discusses the technical report for a new small language model called HY-MT1.5, developed by the Tensen Huanyun team, focusing on its superior performance in translation, especially for languages that are poorly represented in standard training data. The discussion contrasts the efficiency and quality of this small, specialized model against large, commercial models like Google's Gemini 3.0 Pro. ## Detailed Analysis The technical report for HY-MT1.5 details a significant achievement in AI translation, where a small, locally runnable model outperforms massive cloud-based models on specific, challenging tasks. The speakers note that the 1.8-billion parameter HY-MT1.5 model beat the 72-billion parameter Gemini 3.0 Pro model on Chinese-English translation benchmarks, scoring 2.74 to Gemini's 2.55. The core difference lies in the training methodology: HY-MT1.5 uses a five-stage pipeline including instruction tuning and a crucial step involving reverse KL divergence to ensure the student model learns the teacher's entire probability distribution, not just the final answer. This comprehensive approach helps the model handle nuance, context, and cultural appropriateness, which large models often miss, as evidenced by their failure to correctly handle the term 'pilot' in a video game context (translating it as 'aircraft' instead of the correct contextual term). Furthermore, the small model's efficiency allows it to perform real-time, private translation on a local device without internet access, addressing accessibility issues inherent in cloud-based services. ### HY-MT1.5 Performance vs. Giants - HY-MT1.5 (1.8B) outperforms Gemini 3.0 Pro (72B) on specific benchmarks like Chinese-English translation (2.74 vs 2.55) - Excels in low-resource languages like Tibetan and Uger where large models fail due to insufficient data. ### Training Methodology - Utilizes a five-stage pipeline including instruction tuning and reverse KL divergence to accurately mimic the teacher model's full probability distribution, not just the final output. ### Evaluation Rubric - Evaluation goes beyond simple accuracy, focusing on five practical metrics: accuracy, fluency, consistency, cultural appropriateness, and overall utility. ### Limitations of Large Models - Large models often fail on nuanced or context-dependent translations (e.g., 'pilot' meaning a game item vs. an airplane pilot) because they rely on simple metrics or lack targeted fine-tuning. ### Practical Advantages - The small size allows for local, private, real-time translation without internet dependency, directly addressing accessibility issues for low-resource language speakers. ![Screenshot at 00:00: The initial visual features the podcast hosts and the call to action to become a member, overlaid on a waveform graphic.](https://ss.rapidrecap.app/screens/QP36qdYvr1w/00-00-00.jpg) ![Screenshot at 03:38: A visual representation of the 'reverse KL divergence' concept is implied as the speaker discusses forcing the student model to match the teacher's full probability distribution.](https://ss.rapidrecap.app/screens/QP36qdYvr1w/00-03-38.jpg) ![Screenshot at 05:25: The speaker lists the five dimensions used in the evaluation rubric: accuracy, fluency, consistency, cultural appropriateness, and reversibility.](https://ss.rapidrecap.app/screens/QP36qdYvr1w/00-05-25.jpg) ![Screenshot at 08:48: The speaker discusses the pilot example, highlighting the failure of large models to grasp context when translating ambiguous terms.](https://ss.rapidrecap.app/screens/QP36qdYvr1w/00-08-48.jpg) ![Screenshot at 11:55: A comparison of the 1.8B parameter model outperforming the 72B model on specific translation tasks.](https://ss.rapidrecap.app/screens/QP36qdYvr1w/00-11-55.jpg)