Supported presets

Aspose.LLM for .NET bundles ready-to-use presets for several popular open-weight model families. Each preset specifies the model source (Hugging Face repository and file name), context size, chat template, and sampler defaults. Pass a preset to AsposeLLMApi.Create(preset) and the engine downloads the model and any vision projector on first use.

All presets derive from PresetCoreBase (namespace Aspose.LLM.Abstractions.Parameters.Presets). You can use a preset as-is, override any parameter before calling Create, or extend PresetCoreBase for a fully custom model.

Text presets

Preset	Model	Hugging Face source	Default context	Quantization
Qwen25Preset	Qwen 2.5 7B Instruct	bartowski/Qwen2.5-7B-Instruct-GGUF	32 768	Q4_K_M
Qwen3Preset	Qwen 3 8B	bartowski/Qwen_Qwen3-8B-GGUF	32 768	Q4_K_M
Gemma3Preset	Gemma 3	mradermacher/gemma-3-GGUF	8 192	Q4_K_M
Llama32Preset	Llama 3.2 3B Instruct	bartowski/Llama-3.2-3B-Instruct-GGUF	131 072	Q4_K_M
Phi4Preset	Phi 4 mini Instruct	bartowski/Phi-4-mini-Instruct-GGUF	16 384	Q4_K_M
Oss20Preset	gpt-oss 20B (multilingual-reasoner fine-tune)	mradermacher/gpt-oss-20b-multilingual-reasoner-i1-GGUF	131 072	Q4_K_M
DeepSeekCoder2Preset	DeepSeek-Coder-V2-Lite Instruct	lmstudio-community/DeepSeek-Coder-V2-Lite-Instruct-GGUF	163 840	IQ3_M
DeepseekR1Qwen3Preset	DeepSeek-R1 distilled from Qwen 3 8B	lmstudio-community/DeepSeek-R1-0528-Qwen3-8B-GGUF	131 072	Q4_K_M
UnifiedDefaultLlmParameters	Baseline template (no model source)	—	4 096	—

Notes:

• UnifiedDefaultLlmParameters is a conservative CPU-safe template. It sets only context, threads, and sampler defaults; you must set a model source yourself before calling Create.
• Memory requirements scale with context size and quantization. A 7B Q4_K_M model at 32K context needs roughly 6-8 GB of RAM (or GPU memory) for the weights plus KV cache. Longer contexts and larger models need more — see Features for memory guidance.
• All presets ship with preset-specific sampler tuning (temperature, repetition penalty, penalty context size, etc.). Override any field on preset.SamplerParameters before calling Create to change behavior.

Vision presets

Vision presets configure both the base language model and its multimodal projector (mmproj). Pass image bytes via the media parameter of SendMessageAsync or SendMessageToSessionAsync.

Preset	Model	Hugging Face source	mmproj file	Default context	Quantization
Qwen25VL3BPreset	Qwen 2.5 VL 3B Instruct	unsloth/Qwen2.5-VL-3B-Instruct-GGUF	mmproj-F16.gguf	128 000	UD-IQ2_XXS
Qwen3VL2BPreset	Qwen 3 VL 2B Instruct	Qwen/Qwen3-VL-2B-Instruct-GGUF	mmproj-Qwen3VL-2B-Instruct-Q8_0.gguf	262 144	Q4_K_M
Gemma3VisionPreset	Gemma 3 Vision (Latex fine-tune)	mradermacher/Gemma-3-Vision-Latex-GGUF	Gemma-3-Vision-Latex.mmproj-f16.gguf	8 096	Q4_K_M
Ministral3VisionPreset	Ministral 3 8B Instruct (Mistral AI, 2512 release)	mistralai/Ministral-3-8B-Instruct-2512-GGUF	Ministral-3-8B-Instruct-2512-BF16-mmproj.gguf	262 144	Q4_K_M

Supported image formats across all vision presets: JPEG, PNG, BMP, GIF, WebP. Maximum per-attachment size: 50 MB.

Default preset

AsposeLLMApi.GetDefaultPreset() returns a fresh Qwen25Preset instance — useful as a sensible starting point when you do not know which preset to pick. For raw parameter values without a full preset, call await api.GetDefaultParametersAsync().

Picking a preset

If you want…	Try
A balanced general-purpose model	Qwen25Preset or Qwen3Preset
A small, fast model	Llama32Preset (3B) or Phi4Preset (mini)
A long-context model	Llama32Preset or Oss20Preset (131K), DeepSeekCoder2Preset (163K)
A coding-focused model	DeepSeekCoder2Preset
A reasoning-tuned model	DeepseekR1Qwen3Preset or Oss20Preset (multilingual-reasoner)
Image input	Qwen3VL2BPreset (small, very long context) or Qwen25VL3BPreset (3B)
A Mistral family vision model	Ministral3VisionPreset

What’s next

• Presets — preset base class, parameter bags, and override patterns.
• Custom preset — extend or replace a built-in preset.
• Features — full list of capabilities and limits.
• Hello, world! — a minimal runnable example using Qwen25Preset.