PLaMo 2
Japanese-focused large language model series by Preferred Networks
Preferred Networks released PLaMo 2 in September 2025 as a series of large language models focused on Japanese. It uses a hybrid architecture and synthetic data to deliver high performance in Japanese tasks with efficient parameter counts.
- Uses hybrid Samba-based architecture with full attention
- Supports context windows up to 32,000 tokens
- 8B model matches performance of previous 100B version
- Trained on extensive synthetic Japanese corpora
- Optimized for inference with vLLM and quantization
What it is
PLaMo 2 is a collection of large language models designed specifically for Japanese language tasks. Preferred Networks introduced the series in September 2025. The models aim to handle long contexts and provide high-quality responses in Japanese. The largest version contains 100 billion parameters, while an efficient variant uses 8 billion. These systems target instruction following and language fluency within the Japanese domain.
How it came about
The development team at Preferred Networks published a technical report in early September 2025. They revised this document later that month. The work builds on previous efforts to create capable Japanese models. The group sought to address data scarcity issues common in non-English model training. This release represents their latest iteration in focusing computational resources on Japanese linguistic structures.
Inside the model
The system employs a hybrid design based on Samba architecture. It shifts to full attention mechanisms during continual pre-training to support context windows of 32,000 tokens. Training relies heavily on synthetic corpora to compensate for limited natural data. The developers used weight reuse and structured pruning to improve efficiency. This pruning allowed an 8 billion parameter model to match the performance of the earlier 100 billion parameter version. Post-training involved supervised fine-tuning and direct preference optimization, using synthetic Japanese instructions and model merging.
What it can do
The models achieve top results on Japanese benchmarks according to the developers. They outperform other open models of similar size in instruction following and language fluency. The system also demonstrates strong knowledge of Japanese-specific topics. Inference is optimized using vLLM and quantization techniques that preserve accuracy. The 8 billion parameter variant offers comparable utility to the much larger predecessor.
How to use it
Preferred Networks released the technical details via arXiv. The models are optimized for inference using vLLM. Quantization methods are applied to reduce resource needs without significant accuracy loss. The report implies availability for research and community use through standard open channels associated with such publications.
Questions people ask about PLaMo 2
What is the main focus of PLaMo 2?
PLaMo 2 is designed primarily for Japanese language tasks. It focuses on improving instruction following, fluency, and knowledge specific to Japanese, addressing data scarcity through synthetic training corpora.
How does the 8 billion parameter model compare to larger versions?
Through structured pruning and weight reuse, the 8 billion parameter model achieves performance comparable to the previous 100 billion parameter model from the same developer, offering greater efficiency.
What architecture does PLaMo 2 use?
It features a hybrid Samba-based architecture that transitions to full attention during continual pre-training. This design supports long context windows of up to 32,000 tokens.
Good reading and sources
There is no English Wikipedia article on PLaMo 2 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2509.04897. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.