News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Sarashina2

Sarashina2

Open-weight Japanese and English language models by SB Intuitions

SB Intuitions released Sarashina2, a family of large language models with 7 billion, 13 billion and 70 billion parameters. Trained on two trillion tokens of Japanese and English text, these base models use the Llama 2 architecture and are available under an MIT licence for further development.

Share on WhatsApp

  • Trained on 2.1 trillion tokens of text
  • Available in 7B, 13B and 70B sizes
  • Released under an open MIT Licence
  • Uses Llama 2 architecture with RoPE
  • Includes one trillion Japanese tokens

What it is

Sarashina2 is a set of base language models created by SB Intuitions. The series includes versions with 7 billion, 13 billion and 70 billion parameters. These models process text in both Japanese and English. They are built on the Llama 2 architecture and trained on a large corpus of web data. The company provides them as foundational tools for developers to build upon, rather than as finished chat applications.

How it came about

SB Intuitions developed the Sarashina2 family to support Japanese language processing. The training process involved compiling a dataset of approximately 2.1 trillion tokens. This mix included one trillion tokens of cleaned Japanese text from Common Crawl and English documents from SlimPajama. The team excluded the Books3 corpus from the English data due to copyright concerns.

Inside the model

The models use a decoder-only transformer design similar to Llama 2. They employ Rotary Positional Embeddings (RoPE) to handle sequence positions. The vocabulary size is 102,400 tokens across all variants. Tokenization relies on SentencePiece with a unigram language model and byte-fallback, allowing raw sentences to be processed without pre-tokenization. The 70B variant has 80 layers and a hidden dimension of 8,192.

What it can do

As base models, Sarashina2 instances have not been fine-tuned for instruction following. The developers note that they may produce meaningless sequences, inaccuracies, or biased outputs. Users are expected to perform additional tuning to align the models with human preferences and safety standards before deployment in interactive settings.

How to use it

The weights for the 7B, 13B and 70B models are available on Hugging Face. SB Intuitions releases them under the MIT Licence, which permits broad use and modification. Developers can load the models using standard libraries like Transformers and integrate them into their own pipelines.

Safety and testing

The provider states that the models have not undergone alignment tuning. Consequently, they may generate objectionable or biased content. SB Intuitions advises developers to implement their own safety measures and preference tuning before using the models in production environments.

Questions people ask about Sarashina2

Is Sarashina2 ready for chat applications?

No, it is a base model. The developers state it has not been tuned for instructions and may produce biased or inaccurate outputs. Users must fine-tune it for safety and preference alignment.

What languages does Sarashina2 support?

The model supports Japanese and English. The training data included about one trillion tokens of Japanese text from Common Crawl and English documents from SlimPajama.

Can I use Sarashina2 commercially?

Yes, the models are released under the MIT Licence. This permissive licence allows for commercial use, modification and distribution, subject to the terms of the licence.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Sarashina2 yet, so the facts on this page come from the sources listed above: huggingface.co/sbintuitions/sarashina2-70b. 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.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.