Typhoon Translate 1.5 is the next-generation Thai–English translation model that understands not just what to translate, but how you want it translated. Built on the solid foundation of Typhoon Translate, version 1.5 adds instruction-following intelligence while preserving the same lightweight, local-friendly, open-source benefits.
What’s New in 1.5: From High-Quality Translation to Controllable Intelligence
Typhoon Translate 1.5 introduces instruction-aware translation, a major leap that lets users go beyond simple word-for-word translation and specify how the output should look. Whether it’s tone, terminology, formatting, or contextual constraints, the model now understands and follows detailed user instructions with precision.
On our custom Instruction-Following Translation (IF-Translate) benchmark, it achieves roughly 3× higher accuracy than the previous version—demonstrating a dramatic improvement in controllable translation. This new capability gives users fine-grained control over style, terminology, numeric and unit formatting, and even text length, making Typhoon Translate 1.5 not just more accurate, but more adaptable to real-world requirements.
What Stays the Same
🚀 Lightweight: 4B parameters—runs on a regular laptop, no powerful hardware needed.
🔒 Private & Secure: Translates directly on your device; your data stays with you.
🎯 Natural Translation: Delivers fluid, human-like translations on par with leading proprietary models.
🔧 Open Weight: Available now on Hugging Face. Easy to try, fine-tune, or integrate into your own workflow.
The original Typhoon Translate set a new standard for natural Thai–English translation that can run privately on local devices. It excelled at straightforward translation and often matched or exceeded much larger models—without compromising privacy.
However, its primary goal was accuracy, not adaptability. It couldn’t reliably follow stylistic instructions or formatting constraints. Typhoon Translate 1.5 changes that by adding instruction-aware training so users get accurate translations that follow rules.
Methodology
Thai–English translation isn’t a new challenge. For years, datasets like SCB-MT and OPUS have powered most translation systems. But those corpora were built in a pre-LLM era—when quantity often outweighed linguistic fidelity. In today’s landscape of large language models, the benchmark has shifted: quality, contextual accuracy, and adaptability now matter far more than raw data volume.
The first Typhoon Translate was created to meet that new standard through a multi-stage data and training pipeline designed for precision and balance. The process began with data sourcing, collecting diverse, publicly available Thai and English text spanning multiple domains. This was followed by LLM-augmented translation, where several large models—including Gemma-3 27B, Typhoon, and QwQ—were used to produce initial translations. A two-stage labeling process ensured both scale and quality:
- Stage 1 prioritized throughput, generating a wide range of synthetic bilingual data.
- Stage 2 introduced human-in-the-loop validation, where sampled data were reviewed for fluency, accuracy, and tone consistency.
The resulting dataset was then filtered and combined through a data mixture and selection process, producing multiple candidate checkpoints that were benchmarked against both automated metrics and human evaluation. The top-performing checkpoint became the foundation of Typhoon Translate—delivering a strong, natural baseline translation model optimized for local deployment.
Typhoon Translate 1.5 builds directly on that foundation with an extended Stage 3: Instruction Tuning. Beyond learning to translate, the model now learns how to follow instructions while translating. We introduced a new layer of instruction–translation pairs, each encoding explicit, verifiable rules for tone, terminology, formatting, and structural constraints. This additional stage transforms Typhoon Translate from a high-quality translator into a controllable translation system—one that not only produces accurate results but can also adapt its output to match user-defined styles and requirements.
Evaluation Methodology
We benchmarked Typhoon Translate 1.5 to assess its core translation quality and its novel instruction-following capabilities. The evaluation combines standard translation test sets with a custom-built benchmark designed to measure adherence to complex stylistic and formatting constraints.
1. Standard Translation Quality (Thai & English)
A benchmark designed to measure the fundamental accuracy, fluency, and naturalness of translation. This test ensures the model maintains high-quality output on straightforward translation tasks without additional instructions. Our evaluation was conducted on a private test set comprising:
- Thai-to-English Evaluation Data (128 samples)
Thai source texts were selected to ensure balance and coverage from:
-
English-to-Thai Evaluation Data (177 samples)
English source texts were selected to ensure balance and domain diversity from:
2. Instruction-Aware Translation (IF-Translate)
To evaluate the model's primary new feature, we developed a challenging, custom benchmark that tests its ability to follow specific, verifiable rules during translation. This moves beyond simple accuracy to measure fine-grained control over the output. The benchmark includes instructions such as:
- Adjusting Tone & Style: Translating content into a formal or casual tone as requested.
- Enforcing Terminology: Consistently translating specific technical terms or phrases as defined in the rules (e.g., 'ไฮโดรเจนเหลว' must be translated as 'liquid hydrogen').
- Following Formatting Rules: Adhering to specific numeric and unit formatting (e.g., '100 เท่า' → '100×').
- Applying Output Constraints: Respecting limits like word counts or sentence structure.
Our test set included:
- 150 Thai-to-English translation samples.
- 150 English-to-Thai translation samples.
IF-Instruction Translation example:
Evaluation Framework & Results
We used AlpacaEvalHard 2.0, an LLM-as-judge framework, to score the outputs for both benchmarks. This framework is particularly suited for evaluating complex, open-ended instructions and provides a robust measure of quality and instruction adherence.
Figure 1: Comparative Performance on Standard vs. Instruction-Following Translation
Typhoon Translate 1.5 makes a significant leap in instruction-following capabilities. Compared to its predecessor, it achieves a 3.0x improvement on our IF-Translate benchmark, moving from a mean score of 17.65 to 55.10. This establishes it as a leading model for controllable, instruction-aware translation, competitive with top-tier proprietary models like GPT 5.
Figure 2: Detailed Benchmark Scores
Translation Demos
We include structured prompts with verifiable rules, and compare outputs from Typhoon Translate 1.5 vs the original Typhoon Translate.
Example 1: Panel Translation (Tone, names, and style control)
Prompt:
Typhoon Translate 1.5 Result:
Earlier Typhoon Translate Result:
Example 2: Government-Related Translation (Terminology and formal register)
Prompt:
Typhoon Translate 1.5 Result:
Earlier Typhoon Translate Result:
Example 3: Game Context (Do-not-translate lists)
Prompt:
Typhoon Translate 1.5 Result:
Earlier Typhoon Translate Result:
Conclusion
Typhoon Translate 1.5 marks a turning point from a static translation model to a truly controllable translation system. Built on the same open, high-quality foundation as the original Typhoon Translate, it adds a new layer of intelligence that lets users fine-tune every aspect of their output—from tone and terminology to formatting and constraints.
With 3× better instruction accuracy and the same lightweight, privacy-friendly design, Typhoon Translate 1.5 proves that smaller, focused models can deliver world-class performance without compromising control or accessibility.
You can try Typhoon Translate 1.5 right now on Hugging Face or on Ollama.
Limitations
The model was trained on 8192 context length. We recommend limiting input to no more than 8192 tokens for optimal performance.
