Radar 35 advocates for model mix over blanket frontier usage
ThoughtWorks argues that organizations should balance their inference mix between frontier, open-weight, and decision models. This approach aims to optimize costs and performance rather than defaulting to expensive general-purpose models.
The 'Beyond Frontier Models' theme in Radar 35 critiques the practice of using the most powerful frontier models for all tasks, labeling it 'blanket frontier model usage.' This approach is deemed imprudent for cost control and often suboptimal for specific tasks. The radar highlights the rise of specialized and open-weight models that offer comparable performance at lower costs. For example, GLM Legal, a fine-tuned version of the GLM open-source model, outperforms frontier models on legal tasks. Open-weight models like GLM 5.4 and Kimi K3 are noted for achieving coding performance close to frontier labs, making them viable alternatives for many organizations. Additionally, the radar introduces 'decision models,' a category of 'system one' models like JEV (named after Jevons paradox) that make fast, cheap decisions for discrete choices, such as routing tasks or classifying log lines. These models are cheaper and faster than full LLMs, enabling new use cases like reviewing 400,000-line log files for patterns that regex cannot easily capture.