API reference

The complete class-level API reference for Aspose.LLM for .NET is published separately with every class, method, property, and XML documentation comment.

Key types at a glance

The following types form the everyday public surface. The external reference has full signatures, overloads, and remarks.

Aspose.LLM

Type Purpose
AsposeLLMApi The facade class. Create once per process, reuse for all chat operations.
License Applies a license file or stream; exposes the static IsLicensed property.

Aspose.LLM.Abstractions.Parameters

Type Purpose
EngineParameters Engine-wide settings: cache path, debug logging, default threads.
ModelSourceParameters Where to load the model: local path, Aspose ID, or Hugging Face repo.
ModelInferenceParameters Model-load knobs: GPU layers, main GPU, split mode, memory mapping, KV overrides.
ContextParameters llama.cpp context knobs: context size, batch sizes, rope scaling, flash attention, KV dtype.
ChatParameters System prompt, history, max tokens, cache cleanup strategy.
SamplerParameters Full sampler surface: temperature, top-p, top-k, penalties, DRY, XTC, mirostat, seed.
BinaryManagerParameters Native binary release tag, binary cache path, preferred acceleration backend.
MultimodalContextParameters Vision (mtmd) context: GPU use, timings, thread count, verbosity.

Aspose.LLM.Abstractions.Parameters.Presets

Type Purpose
PresetCoreBase Base class holding all nine parameter bags. Extend it for custom presets.
Qwen25Preset, Qwen3Preset, Gemma3Preset, Llama32Preset, Phi4Preset, Oss20Preset, DeepSeekCoder2Preset, DeepseekR1Qwen3Preset, UnifiedDefaultLlmParameters Built-in text presets.
Qwen25VL3BPreset, Qwen3VL2BPreset, Gemma3VisionPreset, Ministral3VisionPreset Built-in vision presets (set both base model and mmproj).

Aspose.LLM.Abstractions.Models

Type Purpose
ChatMessage A single turn in a chat session: role, content, optional media, KV cache metadata.
MediaAttachment Image (or other media) payload with format detection and 50 MB size limit.
MediaFormat Enum: Unknown, JPEG, PNG, BMP, GIF, WebP.
HuggingFaceModel Metadata for a Hugging Face model file (repo ID, filename, size, quantization).
SystemSpec Detected OS, architecture, and available accelerations.
VisionModelInfo / VisionModelSelector Candidate mmproj info and scoring helpers.

Aspose.LLM.Abstractions.Acceleration

Type Purpose
AccelerationType Enum: None, CUDA, HIP, Metal, Vulkan, Kompute, OpenCL, SYCL, AVX512, AVX2, AVX, OpenBLAS, NoAVX.
CudaVersion Detected CUDA major/minor version and installation path.

Aspose.LLM.Abstractions.Interfaces

Type Purpose
ILlamaModel Loaded text model contract.
IMultimodalModel Vision-capable extension of ILlamaModel.
IChatSession Session contract — history, media, KV tracking, response generation.
IModelLoader, IModelFileProvider, IPromptFormatter Extensibility interfaces for custom implementations via DI.

AsposeLLMApi method summary

Method Purpose
Create(preset, logger?) Static factory. Creates the single instance.
DefaultPreset The preset passed to Create.
StartNewChatAsync(preset?, sessionId?) Create a new chat session.
SendMessageAsync(message, media?, preset?, ct) Send to the current session (create one if needed).
SendMessageToSessionAsync(sessionId, message, media?, ct) Send to a specific session.
GetDefaultParametersAsync() Tuple of engine-default inference/context/chat/sampler parameters.
GetDefaultPreset() Returns new Qwen25Preset() as a sensible default.
SaveChatSession(sessionId, filePath?) Serialize a session to disk.
LoadChatSession(filePath) Deserialize a session and set it as current.
ForceCacheCleanup(strategy) Trim the KV cache of the current session.
Dispose() Release native resources and the single-instance guard.

See Chat sessions and Session persistence for detailed semantics.

What’s next