runanywhere-sdks

RunAnywhere Python SDK

On-device AI for Python — LLM, VLM, STT, TTS, embeddings, and voice pipelines run entirely on the host. Inference needs no network; the network is used only to download models.

Current release: 0.20.11

Install

pip install runanywhere==0.20.11

Wheels bundle the compiled native runtime and dependent libraries on Windows, macOS, and Linux. Python 3.9+ and numpy >= 1.21 are required.

Optional extras:

pip install "runanywhere[rag]==0.20.11"      # RAG / document Q&A
pip install "runanywhere[server]==0.20.11"   # OpenAI-compatible local server

Quick start

One initialize() call, then a namespace per modality:

import runanywhere as ra
from runanywhere import LlmOptions

ra.initialize()
print(ra.llm.generate("Explain quantum computing in one sentence.",
                      LlmOptions(model="smollm2-360m")).text)

options.model takes a catalog id (downloaded on first use), a local .gguf path, an https:// URL, or a HuggingFace repo. Once a model is loaded it stays resident, so later calls can leave model out. ra.reset() unloads everything.

RUNANYWHERE_HOME changes where models and state are cached (default ~/.runanywhere).

Text generation

result = ra.llm.generate("What is the capital of France?", LlmOptions(max_output_tokens=32))
print(result.text, result.output_tokens, result.tokens_per_second)

for event in ra.llm.generate_stream("Describe a sunset."):
    if event.is_token:
        print(event.text, end="", flush=True)
    elif event.is_completed:
        r = event.result
        print(f"\n{r.output_tokens} tokens, {r.tokens_per_second:.1f} tok/s")

Multi-turn is a list of messages, not a separate object:

from runanywhere import ChatMessage, Role

ra.llm.generate([
    ChatMessage(Role.SYSTEM, "You are terse."),
    ChatMessage(Role.USER, "Who wrote Hamlet?"),
])

Every verb that blocks has an a-prefixed async twin (agenerate, agenerate_stream, atranscribe, aembed, aquery, …).

Structured output and tools

schema = {
    "type": "object",
    "properties": {"city": {"type": "string"}, "temp_c": {"type": "integer"}},
    "required": ["city", "temp_c"],
}
result = ra.llm.generate_structured("Weather in Paris, as JSON.", schema)
print(result.value, result.valid)

from runanywhere import ToolDefinition

ra.llm.tools.register(
    ToolDefinition(name="get_weather", parameters=schema, description="Current weather"),
    lambda args: {"temp_c": 21},
)
result = ra.llm.generate("Weather in Berlin?")   # the SDK runs the tool and continues
print(result.tool_calls, result.text)

Vision, speech, and voice

from runanywhere import AudioInput, ImageInput, SttOptions, TtsOptions

print(ra.vlm.generate(ImageInput.file("photo.jpg"), "Describe this image.").text)

transcript = ra.stt.transcribe(AudioInput.file("speech.wav"), SttOptions(model="whisper-base"))
audio = ra.tts.synthesize("Hello from RunAnywhere.", TtsOptions(voice="piper-amy"))
speech = ra.vad.detect(AudioInput.file("speech.wav"))

ra.voice.create_session(...) is defined but not available in this SDK: the native bridge binds no voice agent and the package has no microphone or speaker adapter, so it raises SDKException naming the missing symbols. Compose stt → llm → tts yourself. ra.rerank, ra.images, ra.diarization, ra.segmentation and ra.lora raise for the same reason.

Embeddings and RAG

from runanywhere import EmbedOptions, ModelRef, RagDocument

# numpy float32, L2-normalized, returned in input order
vectors = ra.embeddings.embed(["hello world"], EmbedOptions(model="minilm"))

with ra.rag.open(ModelRef("minilm"), ModelRef("qwen2.5-0.5b")) as session:
    session.ingest(RagDocument("Paris is the capital of France."))
    print(session.search("capital of France?"))        # retrieval only
    print(session.query("What is the capital of France?").answer)

Install runanywhere[rag] for RAG support.

Models

Catalog ids (smollm2-360m, qwen2.5-0.5b, smolvlm-256m, minilm, whisper-base, piper-amy, …) download on first load. Anything the generation verbs accept as options.model also works here.

from runanywhere import DownloadEventKind, ModelFilter

for event in ra.models.download("smollm2-360m"):
    if event.kind == DownloadEventKind.PROGRESS:
        print(f"{event.percent}%")

ra.models.list(ModelFilter(downloaded=True))
ra.models.load("smollm2-360m")      # pay the load cost now instead of on first generate
ra.models.state()                   # what is resident + disk used/free

Local OpenAI-compatible server

pip install "runanywhere[server]==0.20.11"
runanywhere serve          # http://127.0.0.1:8000
runanywhere models         # list catalog + download state

Point any OpenAI client at http://localhost:8000/v1. Supports chat/completions (including vision and tools), embeddings, transcriptions, and speech synthesis — all local.

Production defaults restrict clients to catalog model ids, disable server-side image URL fetching (SSRF), and support optional Bearer auth via --api-key or RUNANYWHERE_API_KEY.

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
print(client.chat.completions.create(model="qwen2.5-0.5b",
    messages=[{"role": "user", "content": "Hi"}]).choices[0].message.content)

CLI

pip install runanywhere also installs a runanywhere command with parity to the C++ rcli:

runanywhere run qwen2.5-0.5b "Capital of France?"
runanywhere chat qwen2.5-0.5b
runanywhere stt -i speech.wav
runanywhere tts -t "hello" -o out.wav
runanywhere list --all
runanywhere serve

Use --json for machine-readable output. For the standalone binary, see the RunAnywhere CLI.

Errors

from runanywhere import SDKException

try:
    ra.models.load("missing.gguf")
except SDKException as e:
    print(e.code, e.message, e.recovery_suggestion)

Verbs raise a typed SDKException; nothing returns a result object with a success flag, and nothing hides an error message in a text field. An option the native bridge cannot carry raises rather than being silently ignored.

Notes

One generation at a time per resident model; a concurrent generate raises immediately.
One model per category is resident; asking for a different id in the same category swaps it.
Prompts, responses, audio, and images never leave the host during inference.
Pass api_key + base_url (production) and initialize runs the two-phase control-plane handshake — authenticate, then flush telemetry. HTTP uses a stdlib-urllib transport (no libcurl / third-party client), so the wheel stays dependency-free. With no credentials initialize does no network work.