generate()

generate() is a suspend extension function on RunAnywhere. It returns the full response plus generation metrics once the model has finished.

import ai.runanywhere.proto.v1.LLMGenerationOptions
import com.runanywhere.sdk.public.RunAnywhere
import com.runanywhere.sdk.public.extensions.generate

val result = RunAnywhere.generate(
    prompt = "Write a haiku about Kotlin programming",
    options = LLMGenerationOptions(
        max_tokens = 50,
        temperature = 1.0f,
        top_p = 0.9f,
        stop_sequences = listOf("###"),
    ),
)

println(result.text)
println(result.model_used)
println(result.tokens_generated)
println(result.tokens_per_second)
println(result.generation_time_ms)

result.thinking_content?.let { println("Reasoning: $it") }

options is optional. Omitting it applies LLMGenerationOptions.defaults(). There is a second overload that takes the generated request message, which is what you use for multi-turn history:

suspend fun RunAnywhere.generate(request: RALLMGenerateRequest): RALLMGenerationResult

See Chat for the history form.

LLMGenerationResult

LLMGenerationResult is the Wire-generated ai.runanywhere.proto.v1.LLMGenerationResult, aliased as RALLMGenerationResult. Field names are snake_case.

Field	Type	Description
`text`	`String`	Generated text, with thinking content removed if it was extracted
`thinking_content`	`String?`	Reasoning trace from models that emit one
`input_tokens`	`Int`	Prompt tokens
`tokens_generated`	`Int`	Output tokens
`model_used`	`String`	Model id that served the request
`generation_time_ms`	`Double`	Total wall-clock time
`ttft_ms`	`Double?`	Time to first token, set in streaming mode only
`tokens_per_second`	`Double`	Throughput
`framework`	`String?`	Framework that ran the generation
`finish_reason`	`String`	`"stop"`, `"length"`, `"cancelled"`, `"error"`, or empty

LLMGenerationOptions

LLMGenerationOptions is generated from idl/llm_options.proto. Wire emits a plain class with all-defaulted constructor parameters plus copy(); use named arguments, there is no builder. Defaults come from LLMGenerationOptions.defaults() (com.runanywhere.sdk.generated.convenience), which is generated from the same proto annotations the C header is generated from.

Field	Type	Default	Range
`max_tokens`	`Int`	`100`	`>= 0`
`temperature`	`Float`	`0.8f`	`0.0` to `2.0`
`top_p`	`Float`	`1.0f`	`0.0` to `1.0`
`top_k`	`Int`	`0` (disabled)	`>= 0`
`repetition_penalty`	`Float`	`1.0f` (no penalty)	`>= 0.0`

Other fields on the message: stop_sequences, streaming_enabled, preferred_framework, system_prompt, thinking_pattern, disable_thinking, structured_output, tool_calling, seed, frequency_penalty, presence_penalty, repeat_last_n, min_p, grammar, response_format, echo_prompt, n_threads, and execution_target. On the wire a proto3 zero means “unset, let the engine decide”. defaults() exists for callers who want a populated value; it does not change how an unset field is interpreted. LLMGenerationOptions.validate() throws SDKException with a fieldPath when a value is out of range.

Tuning by task

// Creative
LLMGenerationOptions(max_tokens = 500, temperature = 1.2f, top_p = 0.95f)

// Factual
LLMGenerationOptions(max_tokens = 200, temperature = 0.1f, top_p = 0.5f)

Temperature	Suits
0.0 to 0.3	Code, extraction, factual answers
0.4 to 0.7	General question answering
0.8 to 1.0	Conversational and creative writing
1.1 to 2.0	Brainstorming

Cancelling

import com.runanywhere.sdk.public.extensions.cancelGeneration

val job = lifecycleScope.launch {
    val result = RunAnywhere.generate(longPrompt, options)
}

cancelButton.setOnClickListener {
    lifecycleScope.launch { RunAnywhere.cancelGeneration() }
    job.cancel()
}

cancelGeneration() is a suspend function. It is a no-op when the SDK is not initialized and swallows its own failures, so it is safe to call unconditionally.

Getting Started

Swift SDK

Kotlin SDK

React Native SDK

Flutter SDK

Web SDK

Vibe Coding

LLMGenerationResult

LLMGenerationOptions

Tuning by task

Cancelling

​LLMGenerationResult

​LLMGenerationOptions

​Tuning by task

​Cancelling

LLMGenerationResult

LLMGenerationOptions

Tuning by task

Cancelling