2 min lesson
Picking the base is only the start
Walk through "Picking the base is only the start" in order, then name the proof that tells you it worked.
Step 1 of 2
Picking the base is only the start. Cursor continued pretraining on top of K2.5 to raise its code-domain knowledge before any preference work, and the order of that continued run is itself a design choice.
- 1Short-context continued pretraining. Keep the context window modest at first and pour code-heavy data through it to lift the base's coding knowledge cheaply, before paying the cost of long sequences.
- 2Long-context extension to 256K. Extend the window out to the native 256K (RoPE base/NTK scaling territory) so the model holds a real repo, tool outputs and many turns of history in one session.
- 3SFT on agent-like data. Fine-tune on demonstrations shaped like real Cursor agent traces - tool calls, multi-file edits, test runs - so the model arrives at RL already behaving like an agent, not a chat model.
Cursor runs ONE harness across every model it serves - the same scaffolding for tool calls, context assembly and edit application, whether the model underneath is ComposerCursor's own fast coding model, tuned for the editor and priced well below frontier models; the recommended day-to-day model for executing a plan. Press Enter for the full definition. or a third-party frontier model. Artificial Analysis found that other labs' models score best inside Cursor's harness, which is the striking part: "in some respects it is the product." The lesson for a research candidate is that model quality and harness quality are not separable - a gain you measure has to be attributed to one or the other, and the harness can lift any model you drop into it.
The mid-training data mix shapes everything RL can later reach. If continued pretraining over- or under-weights a language or a code style, RL inherits that bias and can amplify it. A senior answer treats the pipeline as coupled: a data choice three stages back can explain a reward-hacking pattern you see at the end.
When handed an ambiguous quality problem, locate it in the pipeline out loud: "Is this a base-distribution gap (mid-training), a behavior-format gap (SFT) or a reward-optimization gap (RL)?" Diagnosing the stage before proposing a fix shows the end-to-end ownership the role is built around.