Alibaba Cloud published the model-layer design behind Qoder Fast, the coding tier it says cut time to first response from eight seconds to three and credit consumption from 1.1x to 0.8x. The mechanism is a two-stage split. A component the team calls System One performs structured task classification: the application defines the task categories, the model returns one label, and a router uses that label to schedule execution. The classifier is fed only the current request and recent conversation, on the order of a thousand tokens, while the executing model still receives the full context, which for a complex development task can accumulate to hundreds of thousands of tokens. The team reports classification latency in the hundreds of milliseconds, with more than 99% of decisions returned within half a second, measured over more than 1,000 historical task texts. Candidate model priorities are set by offline comparison across more than 20 task types, including implementation, explanation, refactoring, testing and architecture design. Each configuration is judged against acceptance criteria rather than prose quality, for example that a deduplication rewrite preserves first-occurrence order and passes edge-case tests, or that a cross-module refactor keeps existing tests green. Speed is measured end to end, from task start through model calls, tool execution and in-task validation, because the team notes that producing output sooner only shortens time to first response, while longer output or extra tool rounds extend the full task.
Our takeThe design choice is where the reasoning sits: a deliberately small classifier decides, and the expensive model only executes. That inverts the usual instinct of routing with the biggest model available and makes classification accuracy the new single point of failure, since a misread refactor gets an explanation-tier model without anyone noticing. Judging candidates on acceptance criteria and end-to-end time rather than first-token latency is the right frame; the missing number is per-category accuracy of the classifier itself.