At this still-early stage in the evolution of AI, companies are striving to establish differentiators that will determine their long-term relevance and success. And it can be brutally difficult to compete in saturated categories; at Inflection, the company’s CEO and close to 70 other employees jumped ship to Microsoft. Was this a case of one chatbot too many?
Cohere’s latest announcement could see it take one step closer to establishing its key difference, and one that will certainly appeal to the C-suite execs tasked with allocating AI budgets: value for money.
Command R Fine-Tuning
Cohere has announced the general availability on its native platform, as well as Amazon SageMaker, of fine-tuning for Command R, the smallest model in the company’s Retrieval Augmented Generation (RAG)-optimized, R series of AI models. Responding to customer demand for customizable, high-performance models, the newly launched fine-tuning capabilities in Command R enable enterprises to target models to specific business uses by incorporating enterprise-specific language and documents.
Both models in Cohere’s R series, Command R and Command R+, are optimized for tasks such as workflow automation and multilingual support. The models can be easily integrated into business technology architecture and are proven to be both highly efficient and accurate. With fine-tuning, Cohere has amplified these attributes in Command R and provided an LLM that not only performs exceptionally; the company says it does so at a dramatically lower cost than comparable offerings.
Fine-Tuning Cuts Costs
Fine-tuning on Command R enhances performance to a point where the model performs as well, and in some cases better than, larger models but at a much lower cost. Cohere customers can adjust a maximum of five hyperparameters to optimize model performance while benefiting from an LLM that the company says is up to 15X more affordable than comparable models.
Because fine-tuned Command R is smaller than other industry-leading models that offer the same performance, it has a lower inference footprint, making it cheaper to host. “We observe substantial performance and efficacy gains with Cohere’s new models, highlighting consistent effectiveness across disparate applications,” said Erik Bergenholtz, Oracle’s Vice President of AI Strategy and Operations in a Cohere blog post.
“Cohere’s fine-tuned models even outperform out-of-the-box versions of much larger and more expensive models.”
Cohere compared fine-tuning for Command R across a series of popular use cases, including summarization and research and analysis, against GPT-4, GPT-4 Turbo, and Claude 3 Opus.
Cohere found that Command R with fine-tuning showed an average 20 percent performance increase compared to the baseline. In fact, the fine-tuned model demonstrated the highest level of accuracy in summarizations and scored top marks in both scientific and financial Q&As.
The fact that Command R with fine-tuning not only matches, but in many cases outperforms, larger and more expensive models is a big win for Cohere. As companies continue to invest in AI technologies, they will be searching for platforms that offer value without sacrificing performance for cost. With fine-tuning for Command R, Cohere may well have set the bar.
