Contents
Preferred Networks (PFN) has released the weights for PLaMo 3 Translate, an LLM specialized in translation. Available in four sizes—0.8B, 2B, 8B, and 31B—even the smallest 0.8B model outperformed existing models, including some over 30 times its size, on PFN’s benchmarks for Japanese–English translation and translation between Japanese and other languages.
Key Points#
- The weights are available on Hugging Face in BF16, quantized versions, GGUF, and MLX formats.
- The models support 53 languages, surrounding context, glossaries, and specifications for plain or polite style.
- Training data was refined in stages, with smaller models also distilled from the 31B model.
- Personal and noncommercial use is free; commercial use is subject to conditions based on revenue and other factors.
Choose for Your Environment#
The model sizes are 0.8B, 2B, 8B, and 31B. Each comes in formats including BF16, FP8, W4A16, W4AFP8, NVFP4, GGUF, and MLX-4bit. Runtime options include vLLM for GPU environments, llama.cpp for PCs using CPUs or various GPUs, and mlx-lm for Apple silicon Macs.
To get started, check the model card for a format suited to your setup. BF16 suits those prioritizing accuracy, GGUF is useful for running across both CPU and GPU, and MLX-4bit suits memory-efficient use on Macs. Memory requirements and supported environments vary by model and inference settings, so check the model card before downloading.
Building the Data Behind the Smaller Models#
Training consists of two stages of SFT using translation pairs, followed by DPO to favor preferred translations. Models of 8B and below also undergo distillation from the 31B model. The approach first adapts the models to translation using a large dataset, then narrows the data to emphasize quality and content closer to practical use cases.
To synthesize translations, PFN uses “tournament-style data synthesis”: an LLM generates multiple candidates, compares them, and combines their strengths. PFN also uses back-translation, which translates the output back into the original language, and self-distillation, which repeatedly improves training data using the completed translation model. The Phase 1 SFT dataset contains approximately 33.9 billion tokens, and Phase 2 approximately 3.4 billion.
Interpreting the Evaluation Results#
PFN’s evaluation used internal Japanese–English benchmarks, a benchmark for SaaS use cases, WMT24++, and FLORES+. According to the article, the 0.8B model outperformed the translation-specialized models tested on Japanese–English translation and translation between Japanese and other languages. For translation between English and other languages, the 2B model outperformed the comparison models.
These results reflect the benchmark comparisons reported in the article. Results on actual documents vary with subject matter, writing style, proper nouns, HTML handling, and other factors. For long or specialized documents, the most reliable approach is to test the output using text similar to your intended material.
Usage Terms and Demo Caveats#
Private use, such as personal study and hobbies, is free. Noncommercial use, including academic use, has no income or revenue limits and requires no preregistration. Commercial use is free with preregistration through a dedicated form if the user or affiliated company had income or revenue of ¥1 billion or less in the most recent fiscal year. Above ¥1 billion, a separate license agreement is required.
The article also introduces an internal demo running the 0.8B model on a smartphone. However, PFN has not released that app. To use the model on a smartphone, you will need to check compatible formats and runtime environments yourself.
Next Steps#
Choose your intended environment, check the relevant Hugging Face model card for the required format and license terms, then try translating your own text with the 0.8B model.
Source: https://www.preferred.jp/ja/blog/tech/plamo-3-translate-weight-release
Contents
