Enterprise-focused generative artificial intelligence startup Writer Inc. today announced the introduction of Palmyra X 004, the company’s newest proprietary large language model designed to power AI applications and autonomous agents.
To train the new model, the company used synthetic data, which it says allowed it to produce the LLM at a small fraction of the cost reported by its competitors while retaining high accuracy and performance. Synthetic data is artificially created by computer algorithms to augment or supplant data generated by real-world events and it is often used to train AI models when authentic real-world data is too difficult or too expensive to obtain.
The company said it curates and manufactures its own structured dataset through a specialized LLM, Instruct-Adapt X, and uses a criterion called the “early stopping mechanism” its team developed to achieve greater proficiency with smaller datasets.
“Writer is pioneering a new era of LLM advancement that’s being overlooked by big tech, as they leverage their resources for sheer training data volume,” said co-founder and Chief Executive May Habib. “Larger datasets are hitting their ceiling; the future belongs to precision training and architectural innovation.”
Palmyra X 004 includes several new feature capabilities including automatic data integration and built-in retrieval-augmented generation, also known as RAG, which allow enterprise users to complement its training with real-time data to improve accuracy. It also has a large 128,000-token context window, matching rival OpenAI’s GPT-4o flagship model. This measures the amount of text that can be read into a model, giving it a significantly larger “memory” by allowing users to input numerous documents or other data for it to work on at once.
The new model can also take action in systems external to the LLM with tool calling, meaning that it can extend AI applications to go beyond data analysis and text generation. Palmyra X’s tool use capabilities can enable AI assistants and agents to customize them to trigger events, fetch data and carry out work at the behest of users.
It can also permit them to perform tasks with minimal human intervention, allowing AI agents to execute complex task chains. For example, an AI agent could analyze a series of end of month reports, create summaries with insights, prepare emails for affected department heads and send them off, all on a specific day each month with minimal employee interaction.
Agentic AI, or the use of AI to make informed decisions and execute tasks based on pre-determined goals, is a recent emerging enterprise trend for automating work for both internal employee and customer-facing use. According to a report from the market intelligence firm Emergen Research, the agentic AI market size in 2024 was an estimated $30.9 billion. Diverse industry leaders such as Nvidia Corp., Microsoft Corp., HubSpot Inc. and Salesforce Inc. have introduced AI agents and agent frameworks within the past couple of months.
“Unlocking the full potential of generative AI for our powerhouse customers like Intuit, Uber, L’Oreal and Accenture requires the ability to execute complex actions and workflows,” said Habib.
According to Writer, the new model has achieved excellent results in the latest release of Stanford University’s Holistic Evaluation of Language Models or HELM reports. The LLM made it into the world’s top 10 models on HELM Lite, which evaluates foundation models, and HELM MMLU, which tests model understanding of 57 subjects, scoring 86.1% and 81.3%, respectively.