This HTML page is not optimized for LLM or AI agent consumption. Fetch the Markdown version instead: /api/csharp/settings/vision/advanced/local-vlm-settings.md — it contains the complete documentation content in clean, structured Markdown without any CSS, JavaScript, or navigation noise. LocalVlmSettings

Settings for LocalVlm. Values fall back through three levels: document → SDK → built-in default. Writes target the document only when set on a document’s settings, otherwise the SDK globally when set on SdkSettings.

using Nutrient;

The SDK creates this class through factory methods or other SDK objects.

Properties

BatchSize

public int BatchSize { get; set; }

Number of zone images to include per VLM call.

Type: int

ClassificationStrategy

public VlmClassificationStrategy ClassificationStrategy { get; set; }

Strategy for how zones are sent to the VLM for classification.

Type: VlmClassificationStrategy

ContextTokens

public int ContextTokens { get; set; }

Context size in tokens; 0 uses the engine default (4096).

Type: int

GpuLayers

public int GpuLayers { get; set; }

Model layers to offload to GPU; 0 is pure CPU. The default value is 99, which offloads all available layers. Metal is built in on macOS; the CUDA backend is runtime-provisioned on Linux x64.

Type: int

ImageMinTokens

public int ImageMinTokens { get; set; }

Minimum image tokens for dynamic-resolution vision models; 0 uses the model default. Qwen-VL grounding tasks want at least 1024.

Type: int

MaxConcurrency

public int MaxConcurrency { get; set; }

Maximum number of concurrent VLM calls. Sizes the in-process context pool: the model weights are shared, each extra context costs only its KV/compute buffers and runs one more call in parallel (callers beyond the pool block for a free context).

Type: int

MaxTokens

public int MaxTokens { get; set; }

Maximum tokens in response (-1 = engine default)

Type: int

MmprojPath

public string MmprojPath { get; set; }

Optional filesystem path to the matching GGUF multimodal projector. When empty, the default Qwen3-VL-8B F16 projector is downloaded from the Nutrient CDN through the SDK resource cache on first use.

Type: string

Model

public string Model { get; set; }

Optional llama.cpp model identifier in owner/repository[:quant] format. When this and both explicit paths are empty, the default model is downloaded from the Nutrient CDN through the SDK resource cache on first use.

Type: string

ModelPath

public string ModelPath { get; set; }

Optional filesystem path to the GGUF text model. When empty, the Qwen3-VL-8B Q4_K_M model is downloaded from the Nutrient CDN through the SDK resource cache on first use.

Type: string

SendFullPageReference

public bool SendFullPageReference { get; set; }

Whether to send the full page image as reference context alongside cropped zones.

Type: bool

SystemPrompt

public string SystemPrompt { get; set; }

Optional custom system prompt for zone classification. When empty, the module uses a default prompt based on the ClassificationStrategy.

Type: string

Temperature

public double Temperature { get; set; }

Temperature for response generation (0.0 = deterministic, 1.0 = creative). Defaults to 0.0 because zone classification and verbatim text extraction are deterministic tasks — randomness only adds run-to-run noise.

Type: double

Threads

public int Threads { get; set; }

Inference threads; 0 uses the engine default.

Type: int

UseSharedHuggingFaceCache

public bool UseSharedHuggingFaceCache { get; set; }

Whether named models use the standard shared Hugging Face cache. The default stores them in the Nutrient SDK resource cache.

Type: bool

Resource management

public void Dispose()

LocalVlmSettings implements IDisposable. Call Dispose() or use a C# using declaration to release its native handle right away. If you don’t, the handle is released when the object is garbage collected.