Forrester argues that customised, fine-tuned models are now practical for many enterprises. Such models start from open-weight models and are adapted with a company's own data. They are smaller than frontier models and can run closer to where work happens. Forrester names recent examples: Thomson Reuters, Harvey, Travelers, Optimizely and Tech Mahindra. It also notes retreats from general-purpose model building at Databricks and ServiceNow, and a shift at Cohere. Better open-weight models and new post-training tooling are lowering the cost barrier. Forrester still calls the release pace across enterprises a trickle. It advises leaders to compare lifecycle cost against retrieval or agent-based systems. It also suggests testing a self-hosted task-specific model as a failover in resiliency plans. Firms with the largest distinctive data holdings are best placed.
What changed
Building a custom model on a company's own data was too costly and complex for most firms.
What it unlocks
Adapting an open-weight model with proprietary data for a narrow task, and running it on-site or near the edge.
What you need to act on it
- a differentiating proprietary data set
- post-training tooling and skills
Sources