A direct model reasons over the tokens in one prompt. A Recursive Language Model (RLM) keeps the complete input in an external environment and programs against it, bringing selected evidence into model calls.
A traditional LLM loads evidence into its context window before it reasons. An RLM treats the input as a persistent object: the root model writes code to inspect, transform, and reduce it, then uses model attention only where semantic judgment is useful.
Fast and effective when the relevant evidence fits comfortably and the route to an answer is straightforward.
Code does the reduction, while model attention is reserved for the slices that need judgment.
Use hosted context or a connected data source, or run a trusted connector in a private data plane.
The root model receives the question and a description of the working set, then writes code to decide what to inspect.
Code searches and transforms the full working set. Selected slices go to model calls when the task requires judgment.
Receive the answer, status updates, total usage, and a summary of the steps that ran.
A direct call is still the fastest path for a focused request. RLM is for information-dense questions that require exploration, cross-referencing, calculation, and repeated judgment across a much larger working set.
Use it when the relevant input fits, the task is conversational or generative, and latency matters most.
Use it when the job is finding a few likely passages and a retriever can rank the right evidence.
Use it to explore, cross-reference, calculate, and synthesize across records, documents, databases, or codebases.
Published RLM research showed that the execution pattern can push frontier models beyond their native context limits while improving performance on dense reasoning tasks. These results establish the pattern; they are not a guarantee for every ModelRelay workload.
Read the MIT RLM researchManaged RLM is the default onboarding path. Enterprise teams can place the data connector and execution environment in a customer-controlled VPC or infrastructure.
Use the hosted execution API with inline, uploaded, or connected data.
The trusted connector and execution environment run in your boundary while ModelRelay provides policy, model access, and usage attribution.
In both modes: configured root and subcall model providers receive the instructions, schema, and selected data slices needed for their calls. Provider retention and residency depend on the endpoints you choose.
| Direct call | Search / RAG | General agent | ModelRelay RLM | |
|---|---|---|---|---|
| Best at | Focused prompts | Finding passages | Acting across tools | Analyzing a large working set |
| Context strategy | One prompt | Retrieve then generate | Tools, files, or subagents | Code reduction + selected model calls |
| Model choice | Provider-dependent | Implementation-dependent | Framework-dependent | Supported providers by role |
| Product shape | Model API | Search layer | Harness or application | Call it like a model API |
Start hosted, choose your models, and inspect the cost of the complete run.