flutter_llama
library
Classes
ChatChunk
A piece of a streamed reply.
ChatDone
ChatParams
Sampling and chat options.
ChatResult
A whole reply.
ChatTemplate
A chat template override: a llama.cpp built-in name (chatml, llama3, gemma, phi3, mistral-v7, …) or a
template string. null in ChatParams = the model's own.
ChatText
ChatToolCalls
ChatUsage
ClassifyParams
ClassifyResult
DeviceFit
Whether a model fits this device.
DownloadedModelInfo
Информация о скачанной модели
DownloadModelRef
DownloadProgress
Информация о прогрессе загрузки
FileModelRef
FlutterLlama
Main class for interacting with llama.cpp models
FlutterLlamaMultimodal
Мультимодальный класс для работы с llama.cpp моделями
GenerationParams
Параметры для генерации текста
HuggingFaceDownloader
Сервис для скачивания моделей с HuggingFace
HuggingFaceFile
Информация о файле в репозитории HuggingFace
JsonSchemaGrammar
Converts JSON schemas to GBNF rules. Supported: type (string, number, integer, boolean, null, object, array, or a
list of types), enum, const, properties + required (other properties are not allowed; required ones come first in
the declared order, then optional ones may follow), items, minItems / maxItems (≤ 32), anyOf / oneOf, an empty schema (any
JSON value), description (ignored).
LlamaBackend
LlamaConfig
Конфигурация для инициализации модели llama.cpp
LlamaLifecycle
LlamaMessage
One chat message. Assistant messages can carry toolCalls ; tool results carry toolCallId and name .
LlamaResponse
Ответ от модели llama
ModelCatalog
The built-in presets plus those registered by products.
ModelInfo
What the loaded model is (from GGUF meta and the core's facts).
ModelManager
Унифицированный менеджер моделей
ModelRef
A model to check: a GGUF on disk, or one still to download.
MultimodalConfig
Конфигурация для мультимодальной модели
MultimodalInput
Мультимодальный ввод для поддержки текста, изображений и аудио
MultimodalResponse
Мультимодальный ответ от модели
OllamaDownloader
Сервис для работы с Ollama
OllamaModelInfo
Информация о модели в Ollama
OpenModels
Built-in open models (GGUF Q4_K_M; downloadBytes exact as of 2026-10-08; files is the file to download from the Hugging Face repo id).
PresetModel
Информация о предустановленной модели
PresetModels
Список предустановленных моделей
RawDone
The end of a raw generation.
RawFacts
Facts of the loaded model.
RawSampling
Sampling for one raw generation.
SourceStatus
Статус источника модели
ToolCall
A call the model made.
ToolChoice
When the model may call tools.
ToolSpec
A tool the model may call: parameters is a JSON schema.
UnsupportedBackend
The web (and any platform without a native library).
Functions
estimateNeedBytes ({required int fileBytes , int ?kvBytes → int
The bytes a load needs: weights × 1.2 (compute buffers, runtime) plus the f16 KV cache for the context.
labelChoiceGrammar (List <String labels → String
One label out of labels, verbatim.
nativeToolCallGrammar (List <({String name , Map <String Object ? parameters } > tools → String
The native (Qwen / Hermes) tool-call form: <tool_call>\n{"name": "<name>", "arguments": {…}}\n</tool_call>, used
when a tool call is required on a model whose template renders tools natively.
toolCallGrammar (List <({String name , Map <String Object ? parameters } > tools , {bool allowAnswer = true → String
The tool-call grammar: exactly one of {"tool": "<name>", "arguments": <that tool's schema>} or
{"answer": "<text>"} (when allowAnswer).
