How Corti models work

​

S1 is a product, not a model

corti-s1 and corti-s1-mini are not simply relabeled open-weight models. They are production-grade AI products built by Corti. The underlying model parameters are one input; Corti owns the entire operational layer between those parameters and your application. That operational layer includes:

• Model selection: Corti evaluates, selects, and optimizes the strongest available open-weight foundation model parameters based on performance, not vendor loyalty.
• Inference optimization: serving configuration, batching, caching, and hardware utilization are tuned for the best combination of latency, throughput, and cost.
• Quality tuning: models are aligned to produce high-quality outputs on the workloads Corti’s customers rely on, including coding, clinical text, and structured generation.
• Evaluation and guardrails: every release is validated against defined evaluation suites before it becomes the recommended version. Guardrails are implemented and iterated as needed.
• Reliability and scaling: capacity management, failover, monitoring, and observability are part of the product, not left to the customer.
• Continuous improvement: as better parameters, serving techniques, or hardware become available, Corti improves what sits behind the model ID without forcing you to rebuild your application.

​

Standardized model IDs

The model IDs (corti-s1, corti-s1-mini) are stable API contracts. You code against the ID, and Corti continuously improves everything behind it. You never need to track upstream model releases, rename models in your code, or update prompts when the foundation model changes. Think of the model ID as a product name, not a version number. When you specify corti-s1, you get Corti’s frontier-tier model today. When Corti upgrades the foundation model behind that ID, you get the improved model automatically.

​

Foundation models are an implementation choice

Corti S1 starts from some of the strongest open-weight models available. The specific foundation model is selected and optimized by Corti based on performance, and it can change as the frontier moves. Today’s best model will not necessarily be tomorrow’s best model, and Corti absorbs that complexity for you. corti-s1 is currently based on GLM weights and corti-s1-mini on Qwen weights, selected for their respective performance at the frontier and efficient tiers. Corti is not locked into any particular foundation model: if a better architecture or set of weights becomes available, Corti can evaluate it and adopt it behind the same model ID.

​

Could the underlying model change?

Yes. That is a feature, not a weakness. The frontier is moving too quickly to guarantee that today’s strongest underlying weights will remain the strongest six or twelve months from now. Corti’s job is to absorb that change. Customers integrate once. Corti keeps moving the underlying system toward the frontier. Every new S1 release is expected to clear a defined improvement bar before Corti makes it the recommended version. Corti validates new releases across evaluation suites covering accuracy, reasoning, coding, instruction following, tool use, structured outputs, latency, throughput, reliability, safety, and customer workloads.

​

What this means for you

• Pin the ID, not the model: specify corti-s1 or corti-s1-mini in your code. Corti handles upgrades to the underlying foundation model transparently.
• EU vs US: the same model ID may be served by different underlying models in each region. A result from a US project doesn’t tell you about the EU-hosted model behind the same ID, or vice versa. See Regions.
• Version pinning: if you need to pin a specific model version for reproducibility or controlled migration, contact us. Corti supports version pinning so you can choose when to move between releases.
• Sovereign deployment: Corti Models is not limited to a hosted API. The same platform can be deployed in Corti’s sovereign cloud, in a dedicated environment, or on customer-controlled infrastructure including on-premises. Contact us to discuss deployment options.

​

Next steps