Corepass Models

Advancing the frontiers of intelligence.

From deeper reasoning for complex engineering to everyday development and focused tasks, we explore complementary model roles that bring capability and efficiency closer to real work.

Talk to our team

Model

Different strengths. A shared ambition.

Different tasks call for different reasoning depth, response times, and budgets. Our model program explores flagship reasoning for difficult work, balanced capability for everyday development, and specialists for focused tasks.

Horizon

Planned

Flagship reasoning

A flagship reasoning direction for complex engineering and long-running tasks, exploring deeper analysis, coherent decisions, and verification under multiple constraints.

Target tasks
Complex engineering & long-horizon tasks
Capability focus
Depth & coherent decisions

Voyager

In development

Balanced capability

A main coding and reasoning direction for everyday development, testing, and debugging, exploring a balance of task quality, speed, and cost.

Target tasks
Everyday development & iteration
Capability focus
Coding, reasoning & efficiency

Scout

In use within Remy

Lightweight specialists

Scout is a family of specialized models for well-defined, verifiable parts of agent workflows. Its research directions include code retrieval and coding subtasks, with a focus on effective collaboration between specialists and the main agent.

Target tasks
Frequent, focused tasks
Capability focus
Specialization & collaboration

External model APIs are not yet open. Model specifications and pricing will be shared as services become ready.

Development update

Where our models stand.

Follow each model direction through its stage goals, current work, progress so far, and the next questions to test.

Horizon

Flagship reasoning

Stage goal
Deep, coherent, verifiable reasoning for long-horizon engineering.
Current work
Scoping complex reasoning and long-running agent needs.
Progress to date
Results not yet published.
Next steps
Evaluate depth, constraints, and checkable outcomes. Plans to come.

Voyager

Balanced capability

Stage goal
Validate post-training methods for coding agents to inform subsequent Voyager training.
Current work
Building data and evals; testing tool use and feedback learning on small models.
Progress to date
Pipeline connected; a full training step validated. Voyager gains not yet shown.
Next steps
Advance training when results and independent evals support it.

Scout

Lightweight specialists

Stage goal
Balance retrieval precision, recall, and speed on focused tasks.
Current work
Training code retrieval; preparing a coding specialist.
Progress to date
4B retrieval model live in Remy Quick Search; evaluation published.
Next steps
Refine retrieval on independent tasks; prepare coding specialists.

Scout in practice

Progress, backed by evidence.

An applied example from the model program: Scout’s first code-exploration model supports work in Remy. The published evaluation below measures code retrieval and task time.

Scout · First model

Scout

A 4B model that finds the code to change.

Scout handles the retrieval step, so the main agent starts from the right files and line ranges.

4BCode retrievalIn production in Remy

How it works

  1. TaskA change to make
  2. ScoutFinds files and lines
  3. Main agentImplements and verifies

Code-retrieval evaluation

Compared with its base model before RL training, on 848 code tasks.

  • Base model
  • Scout

2.3×

more accurate retrieval

0.077
0.180
Base modelScout
Retrieval F1

29%

faster per task

Base model7.6 s
Scout5.4 s29% saved
Time per task

RSI research

Toward intelligence that improves itself.

Recursive self-improvement (RSI) is our long-term research direction: AI that helps design and evaluate experiments, improving models and agent systems so they can better support the next research cycle.

  1. 01Propose experiments
  2. 02Run and evaluate
  3. 03Improve models & systems
  4. 04Support the next cycle

Advance the next generation of models.

Talk to our team