Pioneering Machine-Native Intelligence.
We build and serve language models as a new class of computational primitives for understanding information, evaluating outcomes, and making decisions, optimized for quality and speed.
Designed to integrate directly into model training and continuous improvement loops, data pipelines, and production workflows, these models power everything from grading and reward modeling to real-time detection, large-scale classification, and structured extraction.
Intelligence Beyond Conversation.
Language models are becoming a new class of computational primitives, enabling systems to understand information, evaluate outcomes, and make decisions.
Unlike general-purpose models designed for open-ended generation, machine-native models are optimized for specific, measurable tasks.
Understand
Turn unstructured information into structured data.
Classification ·Extraction ·Matching
- vendorAcme Ltd
- due2026-03-03
- total1284.00
Evaluate
Assess model outputs, grade agent performance, and generate reward signals.
Grading ·Judging ·Reward Modeling
- verdictpass
- score0.91
Decide
Detect threats, assess risks, and support decisions in real time.
Detection ·Risk Scoring ·Decision-Making
- decisionflag
- risk0.84
Example outputs. Fields and values depend on the model and the schema you give it.
Explore Models →Coming soon
Know Where Models Excel.
No single model is optimal for every task. Some workloads demand exceptional accuracy, others low latency or efficient execution at scale.
Recurrence evaluates models against task-specific benchmarks, revealing their capabilities, limitations, and trade-offs.
Explore model quality, latency, and efficiency to find the right fit for your workload.
Explore Benchmarks →Coming soon
Concept
Longer is better. The pink bar leads its column. Illustrative, not measured.
Inference Built for Your Workload.
Bring models into production with infrastructure optimized for the way they run.
- Real-Time Inference
- Low-latency execution for scoring, detection, and decisions that need immediate results.
- Batch & Async Inference
- Run large-scale classification, extraction, and evaluation asynchronously. Flexible execution windows enable better compute economics.
- Dedicated Inference
- Dedicated capacity and private deployments for workloads requiring predictable performance and greater infrastructure control.
Execution mode
Concept
No window. Each request is answered on its own, as it arrives.
the window you allow when the job runs
- Completion
- Immediate
- Latency
- Per request
- Cost
- Real-time rate
The request runs the moment it arrives, on capacity we keep ready for it, and the answer comes back in the same call.
Shows how the mode and the window change where a job runs and what it costs, relative to real-time. Not a price and not a committed completion time.
Behind every request, Recurrence optimizes model execution, GPU selection, and scheduling to deliver the performance each workload demands.
Model Execution
Task-specific post-training and architectures, so a decision costs a fraction of what a general-purpose LLM would spend on it.
Measured accuracyLower cost per task
Scheduling
Real-time when latency matters. Otherwise a completion window, and the scheduler places the job on idle or spot capacity within it.
Delivered on timeBatch discount
GPU Selection
Chosen for the model's compute profile: memory bandwidth, memory size, throughput. Sourced only from providers that pass our security review. US-only capacity on request.
Security-vettedZero data retention
How much this saves depends on utilization, cold starts and preemption, so we describe the mechanism here and will publish measured figures as they come in from production.
Models That Keep Improving.
The best models are built through continuous evaluation and refinement.
Real-world workloads reveal new challenges, edge cases, and opportunities for improvement. Recurrence combines task-specific evaluation, post-training, and inference optimization to improve model quality and efficiency.
Start with ready-to-use models, then adapt them to your own data, evaluation criteria, and performance requirements.
Concept
Each round is an evaluation, then a change the evaluation justified. Illustrative, not measured.
One Platform. From Research to Production.
Discover and evaluate models, deploy through a unified API, and improve them as your requirements evolve.
Start with a single API call. Scale to millions of requests. Go further with custom models, dedicated infrastructure, and advanced evaluation.
API preview
POST /v1/inference
{
"model": "recurrence/classifier",
"input": "I was charged twice for my order.",
"labels": ["billing", "technical", "other"]
}Response
{
"result": {
"label": "billing"
}
}Proposed API. Endpoints, model ids and response shapes may change before launch.
Our Thinking.
Analysis and opinion on models, inference, and the economics behind them.
October 28, 2026Is hosting an open-source inference engine profitable?Is hosting an open-source inference engine profitable? Public data on DeepSeek V4.1 Flash: H200 never breaks even, B300 breaks even at 35% utilization.
October 25, 2026The case for machine-native intelligenceWhy language models are becoming a new class of computational primitives, why the most capable model is not the right model for every task, and what it takes in evaluation, compute and continuous improvement to serve that.