Amid a quieter period for OpenAI, rival Anthropic has made waves with the launch of its Claude 3 family of large language models (LLMs). Another player to watch in the generative AI space is Israeli startup Deci, which recently announced significant updates to its offerings.
Previously reported in fall 2023 for launching the DeciDiffusion and DeciLM 6B open-source models—fine-tuned versions of Stability's Stable Diffusion 1.5 and Meta's LLaMA 2 7B—Deci has since introduced DeciCoder, a code completion LLM, and the enhanced DeciDiffusion 2.0. Now, the company has unveiled Deci-Nano, a smaller, more efficient closed-source LLM that is currently available only through the Deci Generative AI Development Platform, aimed at enterprises and developers.
Transitioning Away from Open Source?
Deci seems to be shifting towards a more commercial strategy, similar to Mistral’s partnership with Microsoft. This raises questions about the future of open-source AI. Deci’s VP of Marketing, Rachel Salkin, stated via email:
“We remain committed to supporting the open-source community but also recognize the value in closed-source models that enhance accuracy and speed, delivering greater value to our customers.”
Salkin highlighted several open-source models released recently—including DeciLM-6B, DeciLM-7B, and DeciCoder—which continue to see significant downloads, despite their demo spaces being temporarily paused.
Performance at a Competitive Price
If Deci is indeed opting for a commercial path, Deci-Nano is an indicative first step. This model excels in language understanding and reasoning, achieving an impressive 256 tokens in just 4.56 seconds on NVIDIA A100 GPUs.
Deci's blog emphasizes that Deci-Nano outperforms models like Mistral's 7B-Instruct and Google’s Gemma 7B while being exceptionally affordable at $0.10 per million input tokens, compared to OpenAI’s GPT-3.5 Turbo at $0.50 and Claude 3 Haiku at $0.25.
“Deci-Nano embodies our production-oriented approach, focusing on both quality and cost-effectiveness,” said Yonatan Geifman, Deci’s co-founder and CEO. The 8K context window model was developed using Deci’s AutoNAC technology, designed to optimize model efficiency by generating smaller models that closely replicate the functionality of larger ones.
From financial analysis to content creation, Deci-Nano aims to empower businesses to innovate while managing costs effectively.
Deci also provides flexible deployment options, offering either serverless instances for scalability or dedicated instances for enhanced privacy and customization. This versatility allows businesses to adapt their AI solutions according to evolving needs without sacrificing performance.
Launch of a Comprehensive Platform
While much of the recent news centers on Deci-Nano, the introduction of a full Generative AI Platform represents a significant advancement. This platform is described as a “comprehensive solution” tailored to the efficiency and privacy needs of enterprises.
What does it include? According to Deci, users gain access to fine-tunable LLMs, an inference engine, and an AI inference cluster management solution. Deci-Nano is the first proprietary model available through this platform, with plans for more models—some open, some closed—on the horizon.
The inference engine allows users to deploy Deci-Nano according to their needs, whether through Deci’s API, on their virtual private cloud, or on-premises. For clients managing their own virtual private clouds, Deci will supply a containerized model, along with managed inference services in their Kubernetes cluster.
Additionally, the Generative AI Platform offers on-premises deployment options for businesses wanting to keep data in-house. Clients will receive a virtual container housing Deci-Nano and Deci’s Infery software development kit, allowing for integration into their applications.
Pricing details for the Deci Generative AI Platform and its various offerings have not yet been disclosed, but updates will be provided as information becomes available.
Ranking
