PLaMo Translation Model

2026-10-07 更新 / Release

Built with PLaMo. latest は速度と翻訳品質のバランスを確認した Q4_0-ssm-f16（約6.2 GB）です。Mamba の ssm_out 16行列を F16 に保持しています。

latest now points to Q4_0-ssm-f16, a mixed-precision GGUF retaining all 16 Mamba ssm_out matrices in F16. Tested with Ollama 0.40.0 on Apple M1 Max (64 GB).

• Fixed PLaMo 2 rotary-position metadata so stock Ollama initializes 128 rotary dimensions instead of zero. Mamba/attention layer selection is unchanged.
• Corrected BOS handling and explicitly supplied the input language in the translation template. These fixes prevent the missing paragraphs/repetition observed with the previous runtime/template combination.
• Runtime defaults: 32,768-token context, 4,096 output tokens, greedy decoding (temperature 0, repeat_penalty 1, seed 0). The 10M context shown in the catalog comes from upstream model metadata; this release was tested at 32K, not 10M.
• Newly converted weights use pfnet/plamo-2-translate revision cae8da342a3e051ed69f90ce24c23eacff908732 and llama.cpp revision db33d3cb89d8b0d954df66047b13e9cc22df8f10. BF16 and F32 are converted directly from the original safetensors without an F16 intermediate; F32 storage does not add precision beyond the original checkpoint. No new training is implied.
• Calibrated legacy variants retain their original tensor data and receive the same metadata/template fixes; their upstream checkpoint revision was not recorded in the old GGUF. Calibration is preserved rather than replaced with an uncalibrated low-bit conversion.

Available variants

IQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ2_M, IQ3_XXS, Q2_K_S, and IQ3_S were withdrawn after repetition, mistranslated names/terms, omitted content, or loss of the literal “Learn or Die” motto in release validation. TQ1_0 and TQ2_0 could not be imported by the tested Ollama version and were also withdrawn. Previous weights are archived locally for restoration.

Validation and limits

The full English/Japanese passage supplied for this release was used to select and check every retained distinct weight variant. latest scored chrF 73.81, compared with 72.96 for F16, and retained all eight paragraphs, both founders, their love of computer science, and “Learn or Die”. Ollama chat, generate, and raw translation output matched the patched native llama.cpp output after stripping trailing newlines. Four additional short tests on each retained weight variant covered negation, numbers, dates, and Japanese-to-English translation.

These are checks on one selection passage plus four smoke tests, not a general translation-quality benchmark. Wording still differs from the reference. Some retained variants, including F16, omit “Together with You” from the title. There is no guarantee of error-free translation.

The controlled native llama.cpp comparison measured 40.29 generated tokens/s and 11.19 s for the complete passage with Q4_0-ssm-f16, versus 16.67 tokens/s and 25.73 s for F16. These numbers exclude loading and are native llama.cpp measurements, not Ollama performance claims. See the backend PR and reproducible benchmark.

Usage

Refresh an existing download before running:

ollama pull mitmul/plamo-2-translate
ollama run mitmul/plamo-2-translate

The default direction is English → Japanese. For Japanese → English, set the system prompt to the exact language name:

>>> /set system English
>>> 会議は延期されません。明日の午前10時に始まります。

For independent documents, clear the conversation (/clear) or make a fresh API request. The bundled template assumes English input when the target is Japanese, and Japanese input when the target is English; it does not automatically detect the source language.

Example API request:

curl http://localhost:11434/api/chat -d '{
  "model": "mitmul/plamo-2-translate",
  "messages": [
    {"role": "system", "content": "Japanese"},
    {"role": "user", "content": "Do not restart the server until the backup has completed."}
  ],
  "stream": false
}'

For other language pairs, use /api/generate with raw: true and an explicit PLaMo prompt:

<|plamo:op|>dataset
translation
<|plamo:op|>input lang=English
SOURCE TEXT
<|plamo:op|>output lang=Japanese

Replace both language names as needed and keep the final newline. GGUF tokenization adds BOS; do not add a second BOS manually. Other language pairs were not validated in this release.

About the source model

以下のモデル群は、llama.cppを用いて量子化を施したplamo-2-translateのモデルの重みです。

The following model sets are the quantized weights of plamo-2-translate models processed using llama.cpp.

PLaMo翻訳モデルはPreferred Networksによって開発された翻訳向け特化型大規模言語モデルです。 詳しくはブログ記事およびプレスリリースを参照してください。

PLaMo Translation Model is a specialized large-scale language model developed by Preferred Networks for translation tasks. For details, please refer to the blog post and press release.

List of models:

• plamo-2-translate … Post-trained model for translation
• plamo-2-translate-base … Base model for translation
• plamo-2-translate-eval … Pair-wise evaluation model

PLaMo Translation Model is released under PLaMo community license. Please check the following license and agree to this before downloading.

• (EN) https://plamo.preferredai.jp/info/plamo-community-license-en
• (JA) https://plamo.preferredai.jp/info/plamo-community-license-ja

NOTE: This model has NOT been instruction-tuned for chat dialog or other downstream tasks.

For commercial users

Please check the PLaMo community license and contact us via the following form to use commercial purpose.

• (EN/JA) https://forms.gle/mTL8tBLrMYXKNZD56

Bias, Risks, and Limitations

PLaMo Translation Model is a new technology that carries risks with use. Testing conducted to date has been in English and Japanese, and has not covered, nor could it cover all scenarios. For these reasons, as with all LLMs, PLaMo Translation Model’s potential outputs cannot be predicted in advance, and the model may in some instances produce inaccurate, biased or other objectionable responses to user prompts. Therefore, before deploying any applications of PLaMo Translation Model, developers should perform safety testing and tuning tailored to their specific applications of the model.

Acknowledgement

This model is trained under the project, “Research and Development Project of the Enhanced Infrastructures for Post 5G Information and Communication System” (JPNP 20017), subsidized by the New Energy and Industrial Technology Development Organization (NEDO).

AI policies for Preferred Networks, Inc. group

• (EN) https://www.preferred.jp/en/company/aipolicy/
• (JA) https://www.preferred.jp/ja/company/aipolicy/