# Architecture

GPC-1 uses the Qwen3.5-35B-A3B multimodal mixture-of-experts architecture with schema-bound inference. You define the output space; the server scores it and returns typed values.

## Model package

`model/` contains GPC-1's modified backbone weights; `adapter/` contains the matching adapter applied at load time. Use both components from the same release. Qwen identifies the upstream architecture; the packaged weights belong to GPC-1.

The server loads the backbone in BF16 and the adapter in FP32. It uses the existing language-model output projection to score the request's allowed outputs.

| Mode | Output |
| --- | --- |
| Categorical | A selected label and probabilities over your choices |
| Numeric | A grid estimate, probability-weighted mean, and distribution per field |
| Image numeric | Numeric fields conditioned on one image |
| Finite joint | A selected complete record and probabilities over allowed records |

## Numeric ranges

For bounds `[minimum, maximum]`, the 101-point grid is:

`value(i) = minimum + (maximum - minimum) × i / 100`

Numeric fields share one model pass. Each returns a marginal distribution; shared context does not make them a full joint distribution. The probability-weighted mean can fall between grid positions. Angles and other circular quantities need application-specific interpretation.

## Dependent records

Finite-joint mode scores complete records rather than choosing their fields independently. Probabilities are conditional on the supplied records. Compute and memory grow with the number and length of candidates.

## Serving

The runtime verifies packaged asset identities, uses request-local execution, and rejects over-limit inputs without truncation. Its input ceiling is 256K tokens (262,144), including compiled request overhead. See [context configuration](../API.md#context-window) for deployment requirements. Preserve the supplied precision, prompt formatting, and token mapping when reproducing outputs. Scale concurrent traffic with separate replicas.

See the [API guide](../API.md) for request formats and limits.