Fujitsu today announced the launch of Takane, a large language model (LLM) designed for enterprise use in secure private environments. Developed in collaboration with Cohere Inc.,(1) Takane represents a significant leap forward in generative AI capabilities, offering world-class Japanese language capabilities.
JGLUE(4) Score List. Image Credit: Fujitsu
Fujitsu will integrate this new LLM into its generative AI services on Fujitsu Kozuchi and offer it through Fujitsu Data Intelligence PaaS (DI PaaS), an all-in-one operation platform that is part of Fujitsu Uvance, Fujitsu’s portfolio of solutions addressing cross-industry societal challenges. Takane will be available globally starting September 30, 2024.
Takane, which has achieved world-leading results on the Japanese General Language Understanding Evaluation (JGLUE) benchmark(2), is designed for enterprise use in a secure private environment. Fujitsu will offer Takane alongside its generative AI framework for enterprises which comprises knowledge graph extended RAG technology for referencing large-scale text and monitoring technology for generative AI that ensures output is compliant with laws, regulations, and corporate rules. With this comprehensive approach, Fujitsu aims to create an LLM that supports customers' business transformation.
With the launch of Takane, Fujitsu elevates its generative AI offerings on Fujitsu Kozuchi with a high-precision LLM tailored for secure private environments. This strategic move aligns with Fujitsu's vision for seamlessly integrating generative AI into business operations.
Takane will also be provided under an initiative to deliver total support for customers' generative AI journeys by combining the LLM with its Uvance Wayfinders consulting service and broader Fujitsu Uvance offerings. With this comprehensive approach, Fujitsu aims to empower customers to unlock new value and address societal issues by enhancing productivity, creativity, and innovation in their business operations.
Vivek Mahajan, Corporate Vice President, CTO, CPO, Fujitsu Limited, comments:
“Takane represents a key part of Fujitsu’s AI strategy, which was unveiled last year. Takane will empower companies in industries that demand the highest level of security to harness the power of generative AI. Fujitsu is committed to creating new value for businesses by developing cutting-edge AI technologies. This commitment fuels our three growth drivers: modernization, Fujitsu Uvance, and consulting, which are all powered by our innovative technology.”
Yoshinami Takahashi, Corporate Vice President, COO, Fujitsu Limited, comments:
“Fujitsu Uvance supports our customers’ business growth and the resolution of societal challenges through advanced decision-making supported by data and AI. We are excited to offer Takane, a state-of-the-art LLM specifically optimized for the Japanese language to a wider audience via our DI PaaS platform. We are dedicated to support customer’s business transformation by bringing the most advanced AI to market, not only through our own innovations, but also by collaborating with our global partners.”
Aidan Gomez, Co-founder and CEO, Cohere Inc., comments:
"We are very excited to bring Takane’s advanced Japanese LLMs to global enterprises. Our partnership with Fujitsu accelerates AI adoption in this critically important market by offering secure, performant AI designed specifically for business use across Japanese and other languages."
Growing demand for specialized LLMs optimized for Japanese business needs
The adoption of generative AI is increasing globally across industries including marketing and customer service. However, general-purpose LLMs are often delivered through public cloud services and are therefore not suitable for use in industries handling confidential data and for tasks requiring compliance with laws, regulations, and industry rules. Furthermore, characteristics of the Japanese language, including the mixed use of multiple character types, omitted subjects, and honorific expressions also pose significant hurdles for general-purpose LLMs, limiting their accuracy and reliability in business contexts. A highly accurate Japanese LLM is especially important for sectors including government, finance, healthcare, and law, where even minor language errors can have serious consequences.
Takane Features
1. Strong Japanese Language Capabilities
Takane is based on Cohere's LLM Command R+ and combines Fujitsu’s extensive knowledge in developing Japanese-specialized LLMs with Cohere's expertise in creating task-specific language models. Takane will be offered exclusively by Fujitsu around the world. Through additional training and fine-tuning to enhance Japanese language capabilities, Takane has achieved industry-leading performance on the JGLUE benchmark for Japanese language understanding, surpassing competitors (3) in natural language inference (JNLI), reading comprehension (JSQuAD) tasks, and semantic understanding and syntactic analysis as demonstrated by its top ranking on the Nejumi LLM leaderboard3. Takane further includes Command R+’s multilingual support (covering 10 languages) and its ability to automate business processes.
2. Secure Private Environment
Takane is a medium-sized LLM that can be used in a secure, private environment. This allows for its safe use in industries where data breaches are a concern, such as finance, manufacturing, and the security sectors, which all handle sensitive data.
3. Specialization Through Fine-Tuning
Takane can be further specialized for customer operations through fine-tuning and customization using company data. Cohere's RAG technology in combination with Fujitsu’s knowledge graph extended RAG and generative AI auditing technologies further facilitate compliance with laws and regulations, as well as industry and corporate rules. Fine-tuning and RAG technology make it possible to use Takane even in industries with specialized terminology and frequent regulatory revisions, such as the financial industry.
Fujitsu will offer Takane through its DI PaaS platform, thereby creating a business application that combines data and AI. DI PaaS empowers organizations to unify vast, disparate data sources, both internal and external, and glean meaningful insights. By enabling seamless connection and analysis of data across departments and industries, DI PaaS helps customers to unlock new knowledge and solutions, and fosters data-driven innovation across organizations.
Fujitsu will unlock increased productivity and creativity and continue supporting customers’ business transformation through AI/data integration.
Scheme for offering Takane. Image Credit: Fujitsu
Takefumi Yamamoto, Operating Officer, Deputy Group Chief Information Officer, Mizuho Financial Group, Inc., comments:
“In fiscal 2023, we conducted joint trials with Fujitsu to explore the potential of generative AI in system development and maintenance. Since fiscal 2024, we have been using Fujitsu Kozuchi to enhance system quality by streamlining parts of the system development process. Now, with the launch of the enterprise LLM Takane, we are excited to unlock even greater possibilities. By leveraging knowledge graph extension RAG and other cutting-edge technologies, we will be able to utilize our internal knowledge much more efficiently. We’re committed to continuously improving the quality and resilience of our system development and maintenance processes, and we believe ‘Takane’ will be a valuable tool in achieving this goal.”
Shoji Tanaka, Ph.D, Senior General Manager, Corporate AI Strategy Division, Deputy Senior General Manager, DX Innovation Center, Mitsubishi Electric Corporation, comments:
“Takane seamlessly merges the world-class technologies of Fujitsu and Cohere. This powerful combination opens up a whole new realm of possibilities for the generative AI market in Japan, and I am excited to see what the future holds for both companies.”
Yuki Shuto, Partner, Chief Strategy & Innovation Officer, Asia Pacific Leader of Technology, Media and Telecommunications, Deloitte Tohmatsu Consulting LLC, comments:
"Generative AI has moved beyond proof-of-concept in numerous organizations, transitioning into a phase of full-scale integration across products, services, and operations. The key to success will be secure, reliable LLMs that enable organizations to confidently share and leverage their data. Takane, with its enhanced Japanese language capabilities, holds immense potential to accelerate transformation within Japanese businesses and government agencies. Deloitte Tohmatsu, in collaboration with Fujitsu, is committed to unlocking the full potential of generative AI and contributing to the growth of Japan's economy and society.”
Takafumi Yano, CEO, RUTILEA, Inc., comments:
“Under our slogan ‘AI Made Easy,’ we are committed to providing highly accurate AI services that will enable the integration of AI into all business processes. We are currently pursuing vertical AI initiatives in specific industries, including central government, local governments, the automotive industry, the pharmaceutical industry, and power procurement optimization. In 2024, we will launch an AI development platform business with the establishment of a GPU data center in Okuma Town, Futaba District, Fukushima Prefecture, Japan. 'Takane' holds the potential to significantly accelerate the adoption of AI in various vertical industries across Japan. We look forward to the day when we can contribute to its development and utilization."
- Cohere Inc: Headquarters: Toronto, Ontario, Canada; San Francisco, California, USA; Co-founder and CEO: Aidan Gomez
- Achieved world-leading results on the Japanese General Language Understanding Evaluation (JGLUE) benchmark: Takane outperformed other LLMs on JGLUE, including large-scale, general-purpose LLMs provided through cloud services. For JNLI and JCoLA, there was uncertainty in the correct data (ground truth), so the correct data were corrected by multiple annotators and measured as reference values (as measured by Fujitsu and Cohere in September 2024).
- Surpassing competitors: On the Nejumi LLM Leaderboard 3, which evaluates the Japanese language capabilities of LLM models, Takane achieved the highest performance in the semantic understanding category with a score of 0.862 and in the syntactic analysis category with a score of 0.773 (as measured by Fujitsu and Cohere in September 2024).
- JGLUE: Japanese version of GLUE (General Language Understanding Evaluation) (JGLUE: Japanese General Language Understanding Evaluation; Kentaro Kurihara (Waseda University), Daisuke Kawahara (Waseda University), Tomohide Shibata, 28th Annual Meeting of the Language Processing Society of Japan (NLP2022))
- JSTS: Task to estimate the semantic similarity between a pair of sentences.
- JCoLA: Task for syntactic evaluation of Japanese language models
- JNLI: Task for natural language inference that recognizes the inference relationship between a premise sentence and a hypothesis sentence.
- JCommonsenseQA: Task to assess commonsense reasoning ability.
- JSQuAD: Machine reading comprehension task to evaluate ability to read a document and answer questions about it.