1 min lesson
Base-model selection as a research decision
Name the parts in "Base-model selection as a research decision" and give the practical job of each one.
Step 1 of 2
Base-model selection as a research decisionComposer 2 chose Kimi K2.5
Cursor has stated Composer 2Cursor's in-house agentic coding model: frontier-level coding quality at high speed and low cost, built as a software-engineering specialist rather than a general-purpose model. Press Enter for the full definition. was built on Kimi K2.5The open base model Cursor continued-trained into Composer 2 (1T parameters, 32B active, 256K context), chosen mainly for how well it fit Cursor's serving infrastructure. Press Enter for the full definition., chosen on concrete internal criteria rather than a leaderboard number. That's the texture to bring: base selection is an experiment, not a vibe.
- Internal-codebase perplexity
- How well the base predicts Cursor's own real code, not a generic benchmark. A direct proxy for fit to the target distribution.
- Coding knowledge
- Breadth and depth of programming competence already baked in, so post-training builds on a strong prior rather than fixing fundamentals.
- State tracking
- Ability to follow program/agent state across many steps - the prerequisite for long-horizon agent-assisted editing.
Each criterion maps to a downstream need: distribution fit, prior competence and long-horizon coherence.
Kimi K2.5The open base model Cursor continued-trained into Composer 2 (1T parameters, 32B active, 256K context), chosen mainly for how well it fit Cursor's serving infrastructure. Press Enter for the full definition. was also picked because its architecture fit Cursor's infrastructure. The shapes that matter at training and serving time - the MoE layout, the attention scheme, the native context window - were already close to what the 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. stack runs, so adopting it meant less infra surgery before the interesting work started.
Learn more
Advanced table
Chosen for infrastructure fit, not a leaderboard rank
- Spec
- Total parameters
- Value
- 1T
- Why it mattered to Cursor
- Large MoE capacity to learn from a broad code corpus
- Spec
- Active per token
- Value
- 32B
- Why it mattered to Cursor
- Only a fraction fires per token, so serving cost stays affordable
- Spec
- Layers
- Value
- 61
- Why it mattered to Cursor
- Depth for long-horizon state tracking across an agent session
- Spec
- Native context
- Value
- 256K
- Why it mattered to Cursor
- Long agent sessions fit without aggressive truncation
- Spec
- Attention
- Value
- Multi-head latent (MLA)
- Why it mattered to Cursor
- Compresses the KV cache, easing the inference-memory bottleneck
| Spec | Value | Why it mattered to Cursor |
|---|---|---|
| Total parameters | 1T | Large MoE capacity to learn from a broad code corpus |
| Active per token | 32B | Only a fraction fires per token, so serving cost stays affordable |
| Layers | 61 | Depth for long-horizon state tracking across an agent session |
| Native context | 256K | Long agent sessions fit without aggressive truncation |
| Attention | Multi-head latent (MLA) | Compresses the KV cache, easing the inference-memory bottleneck |
Kimi K2.5 base specs Cursor started Composer 2 from - chosen for infrastructure fit, not a leaderboard rank.