In a casual chat, a student told their mentor, 'Kimi's model is fine, but its framework feels lacking.' The mentor paused, realizing the student had intuitively split AI products into three distinct layers: the model itself, the surrounding framework, and the underlying compute resources. Most people lump these together when criticizing AI tools, leading to vague feedback and misaligned expectations. The model refers to the core neural network and its training; the framework encompasses the user interface, APIs, and integration logic; compute covers the hardware and infrastructure that powers inference. Understanding these layers helps developers identify where bottlenecks truly lie—whether it's a weak model, a clunky UI, or insufficient GPU capacity. For product teams, this clarity enables more targeted improvements and honest communication with users. The anecdote underscores a broader lesson: sometimes the most insightful technical observations come from non-experts who describe what they experience in plain language.
A student's offhand remark about 'the model is fine, but the framework isn't' exposes a common conflation in AI product discussions. This piece unpacks the three distinct layers—model, framework, and compute—and why separating them matters for clearer communication and better product decisions.