Preferred Networks opens PLaMo 3 Translate weights in four sizes, 0.8B to 31B
Preferred Networks released PLaMo 3 Translate, a translation-specialized LLM, as open weights on November 5 in four sizes: 0.8B, 2B, 8B and 31B. The company claims world-class translation performance and says even the smallest model outperforms Google's TranslateGemma and Tencent's Hy-MT2.
Preferred Networks (PFN) released PLaMo 3 Translate, a translation-specialized large language model, as open weights on November 5, 2026. The model comes in four sizes: 0.8B, 2B, 8B and 31B parameters. PFN claims "world-class" translation performance. The announcement was reported by ITmedia AI+ in a Japanese-language article.
Sizes and performance claims
The four variants are 0.8B, 2B, 8B and 31B. According to PFN, translation quality between Japanese and English, and across multiple languages, exceeds that of Google's translation-specialized LLM TranslateGemma and Tencent's Hy-MT2, even at the smallest model size. These are company claims. Specific benchmark scores are not given in the ITmedia article, which links to a PFN press release containing the performance evaluation.
PFN also says the models can run on-device, including on smartphones and laptops.
Availability and license
The weights can be downloaded from Hugging Face. They are released under the PLaMo Community License, which PFN defines. Under that license, companies and individuals whose revenue or sales in the most recent fiscal year was under 1 billion yen can use the models commercially.
Key facts
- Developer: Preferred Networks
- Model: PLaMo 3 Translate
- Sizes: 0.8B, 2B, 8B, 31B parameters
- Release date: November 5, 2026
- Distribution: open weights on Hugging Face
- License: PLaMo Community License; commercial use permitted for entities with prior-fiscal-year revenue under 1 billion yen
- Compared against (per PFN): Google TranslateGemma, Tencent Hy-MT2
- Context window, training cutoff: not stated in the ITmedia article
What this means
The 0.8B to 31B range gives users a choice between models small enough for phones and laptops and a larger model for heavier workloads. PFN's claim that the smallest variant beats TranslateGemma and Hy-MT2 on Japanese, English and multilingual translation is its own, and it rests on PFN's evaluation in the linked press release. The commercial-use condition is tied to revenue: organizations under the 1 billion yen threshold can use the weights commercially under the PLaMo Community License, so anyone planning to deploy it should check the license terms against their own revenue.
Related Articles
China Merchants Lion Rock Lab releases Lion-ER-35B-A3B, an open-weight embodied reasoning model with 3B active parameter
China Merchants Lion Rock AI Lab released Lion-ER-35B-A3B, a mixture-of-experts embodied reasoning vision-language model with about 35B total and about 3B active parameters per token. The lab says it beats its base model on all ten public benchmarks tested, with an equal-weighted average of 65.29. Weights are open on Hugging Face.
Sarvam AI releases Prism, a 105B-based decision model that claims 44 ms responses and a 71.8 average score
Sarvam AI has launched Prism, a System 1 decision model built as a LoRA fine-tune of Sarvam 105B. The company claims 44 ms response times and a 71.8 average score across three benchmarks. Weights are on Hugging Face under Apache 2.0.
Microsoft releases Mage-VL-Cowork, a 4B computer-use agent scoring 66.0% on OSWorld-Verified
Microsoft has published Mage-VL-Cowork on Hugging Face, a 4B-parameter vision-language policy that combines GUI control and code/terminal execution in one checkpoint. According to the model card, it reaches 66.0% on OSWorld-Verified under a 50-step budget, the best among 4B models. It is released under Apache-2.0.
Moxin Technology launches MoWorld 2.0, a 4D world model trained and deployed natively on Huawei Ascend
Moxin Technology unveiled MoWorld 2.0 at a World Internet Conference forum in Wuzhen on November 4. The 4D world model was trained and deployed natively on Huawei Ascend hardware and adds physics modeling, with output as 4D Gaussian assets. Moxin claims inference cost can fall to as little as 30% of a same-scale GPU setup.
Comments
Loading...