Skip to content
Coming soonExplore the public preview. Connected model, compute and payment services are in development.Service status
SAVRN
Search Contact SAVRN
SAVRN Cloud

Models and inference

Streaming, tools and limits

A request can ask for structured extraction or code checks without proving that a runtime enforces a schema or executes a test. Review the separate evidence required for each feature, the proposed candidate limits and the browser simulation boundary before interpreting a familiar control as supported connected service behavior.

Download Markdown

SAVRN Cloud · Coming soon · Public preview · Browser-local simulation · No live compute or payments

Distinguish a request from a supported feature

A guided starter expresses what a researcher wants. It does not qualify that capability in a model. Asking for JSON in Structured extraction is not proof of schema enforcement. Asking for checks in Research code does not execute those checks. The Evidence review starter stays within supplied text and does not invoke a browsing tool.

The Playground runs a deterministic browser fixture. Any visible chunking or apparent response timing demonstrates interface behavior, not model streaming. A receipt should retain the intended candidate separately from the fictional fixture that supplied the sample.

Read the proposed candidate bounds

The prepared Qwen3.5-9B Q6_K and Qwen3.8-27B Q5_K_M configurations each propose 8,192 total context tokens and 1,024 output tokens for a pilot. These are planned qualification limits. They are neither measured service guarantees nor the maximum context advertised for the base models. Demonstration controls have their own sample range and must not be treated as accepted production settings.

Qualify each behavior separately

Capability Evidence needed before release
Text completion Accepted input, valid output, exact runtime identity and metering.
Streaming Terminal events, interruption handling and reconciled usage.
Structured output Supported schema subset, validation and explicit failure behavior.
Tool calls Defined arguments, authorization and a separately controlled executor.
Context and output limits Boundary tests, rejection behavior and resource limits.

No optional feature becomes available merely because a base model supports it elsewhere. Image inputs, tools and embeddings also require their own interface and qualification work.

Keep the execution boundary clear

Connected services are Coming soon. Do not supply credentials, private research material or executable production instructions to test a feature in this preview. Source review and sample preparation can proceed with public or invented inputs. A future release must document concurrency, deadlines, cancellation, retry behavior and supported client versions; see API publication.

Build toward the work that matters.

Tell SAVRN what your institution needs to run, who reviews the results and where its data must stay. That workload defines the next service to qualify.

Discuss an AI project