PLaMo 3 Translate: Four model sizes, local translation

1. Translation across 53 languages on devices and in-house servers

PFN released the model weights on Hugging Face on November 5, 2026. The four sizes have 800 million, 2 billion, 8 billion and 31 billion parameters, the numerical values adjusted during training. They support translation across 53 languages, including Japanese and English. Users can choose a size based on quality, response speed, throughput—the amount processed in a given time—and hardware.

Running on in-house servers or devices keeps source texts and translations of contracts, R&D materials and customer support documents off external servers. Users keep information management internal but must configure the runtime environment and select a deployment format suited to their hardware.

Deployment options include standard GPU weights and quantized formats that reduce size by lowering numerical precision. GGUF for CPUs and GPUs and MLX for Apple silicon are also available.

2. Data refinement improves smaller models

According to the company, the smallest 800-million-parameter model outperformed comparison models from Google DeepMind's TranslateGemma and Tencent's Hy-MT2 in Japanese–English and Japanese–multilingual translation. The comparisons included models with more than 30 times as many parameters. The company also reported that its 2-billion-parameter model outperformed the comparison models in English–multilingual translation.

The Japanese–English comparison used an average of scores from two in-house evaluations and an international translation evaluation. The multilingual comparison selected 15 supported languages and tested translation in both directions between Japanese or English and the other languages. Korean was among the evaluation languages.

PFN refined training translations by combining the strengths of multiple candidates and also used natural human-written sentences as reference translations. It said smaller models also used distillation, learning to follow the completed 31B model's predictions. This method trains smaller models on a larger model's output to reduce size while improving quality.

3. Check translation scores alongside file size reductions

Four 31B deployment formats were compared using the company's model card, which explains the model and its usage. BF16 is the unquantized baseline. Translation scores measure similarity to reference translations; higher is better within each column.

Table: File sizes and translation scores by 31B deployment format

Conditions: 31B model; deployment and quantization formats are listed by row. PFTB evaluates Japanese–English translation using bidirectional charsim, a character-level translation similarity metric. FLORES+ is a multilingual translation evaluation dataset. Its score here is the mean bidirectional chrF++ between Japanese and the other 14 of the 15 evaluation languages. chrF++ measures overlap in character and word fragments; Japanese output uses polite style.

Units: Weight file size in GB; translation quality in each metric's score

The two middle formats in the table have smaller files than the baseline and similar scores on both translation metrics. The smallest format scores lower on both. This shows the trade-off between storage savings and translation quality. Besides fitting hardware capacity, translations of actual work documents are also a selection criterion.

4. Run directly or connect to a translation server

Developers can run the model directly using the model card examples or start a vLLM server and send API requests. For service integration, the recommended approach is to build model-specific prompts directly and send them through the Completions API.

The server example sets a combined input and output limit of 16,384 tokens, adjustable to available memory. The maximum combined length during training was 32,768 tokens. When splitting documents, both source text and translation must fit within the runtime environment's limit.

Below is the chat API request example from the model card.

curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "plamo-3-translate",
    "messages": [{"role": "user", "content": "こんにちは。"}],
    "chat_template_kwargs": {"source_lang": "ja", "target_lang": "en"},
    "temperature": 0,
    "max_tokens": 512,
    "stop": ["<|plamo:tag|>", "<|plamo:bos|>"]
  }'

The target language is required; the source language, reference context and glossary are optional. Japanese output can use polite or plain style. These settings are passed as template arguments.

5. Check revenue thresholds and registration for commercial use

Noncommercial use, including personal and academic use, is free and requires no advance registration. Commercial use is free after advance registration if the user or its affiliates had income or revenue of 1 billion yen or less in the latest fiscal year. Above that threshold, a separate commercial license agreement is required.

In a correction on the release date, PFN changed the free-use revenue threshold from 'less than 1 billion yen' to '1 billion yen or less.' Users exactly at the threshold also qualify for free commercial use after registration.

The release of PLaMo 3 Translate's weights expands self-hosted options, letting users choose translation quality, hardware capacity and information management together.