generate()

RunAnywhere.llm.generate runs a single non-streaming completion and returns the text with token counts and timings.

Basic usage

let result = try await RunAnywhere.llm.generate(prompt: "Name three colours")
print(result.text)

With options:

let result = try await RunAnywhere.llm.generate(
    prompt: "Explain quantum computing in simple terms",
    options: LlmOptions(
        model: "qwen3-0.6b",
        maxOutputTokens: 200,
        temperature: 0.7
    )
)

print(result.text)
print("\(result.tokensPerSecond) tok/s")

Generation auto-loads whatever it needs. When options.model names a model that is not on disk, it downloads it first.

Signatures

func generate(prompt: String, options: LlmOptions? = nil) async throws -> GenerationResult
func generate(messages: [ChatMessage], options: LlmOptions? = nil) async throws -> GenerationResult

The messages: form is covered in Chat.

LlmOptions

Field	Type	Default
`model`	`String?`	none; uses the loaded model
`maxOutputTokens`	`Int`	proto default
`temperature`	`Float`	proto default
`topP`	`Float`	proto default
`topK`	`Int?`	none
`minP`	`Float?`	none
`frequencyPenalty`	`Float?`	none
`presencePenalty`	`Float?`	none
`repetitionPenalty`	`Float?`	none
`seed`	`Int?`	none
`stopSequences`	`[String]`	`[]`
`systemPrompt`	`String?`	none
`reasoning`	`ReasoningOptions?`	none
`structuredOutput`	`StructuredOutput?`	none
`tools`	`[ToolDefinition]`	`[]`
`toolChoice`	`ToolChoice`	`.auto`
`maxToolCalls`	`Int`	`5`
`autoExecute`	`Bool`	`true`

Sampling defaults come from RALLMGenerationOptions.defaults() rather than being hardcoded, so they track the C++ core.

GenerationResult

Field	Meaning
`text`	Generated text, with thinking content removed
`tokensPerSecond`	Throughput

Reasoning models return their thinking channel separately when ReasoningOptions(includeInOutput: true) is set, rather than inline in text.

Reasoning

let result = try await RunAnywhere.llm.generate(
    prompt: "If a train leaves at 3pm going 60mph, when does it arrive 150 miles away?",
    options: LlmOptions(
        reasoning: ReasoningOptions(mode: .on, includeInOutput: true)
    )
)

ReasoningMode is .on by default. The C++ core owns tag parsing and /no_think directives; you toggle the mode and render what comes back.

Structured output

let result = try await RunAnywhere.llm.generateStructured(
    prompt: "Extract the name and age: Dana is 34.",
    schema: schema,
    mode: .validationOnly
)

let person = try result.decode(Person.self)

mode: .constrained throws SDKException today. Engine-level constrained decoding is not wired in yet. Use .validationOnly (generate, then validate) or .repair (validate, then retry once with a repair instruction).

Errors

Throws SDKException when no model can be loaded or generation fails. See Error handling.

Getting Started

Swift SDK

Kotlin SDK

React Native SDK

Flutter SDK

Web SDK

Electron SDK

Vibe Coding

Basic usage

Signatures

LlmOptions

GenerationResult

Reasoning

Structured output

Errors

​Basic usage

​Signatures

​LlmOptions

​GenerationResult

​Reasoning

​Structured output

​Errors

Basic usage

Signatures

LlmOptions

GenerationResult

Reasoning

Structured output

Errors