orchard 2026.5.10

Python client for Orchard, a compute platform for Apple Silicon

Orchard

Standalone local inference for Apple Silicon, from Python.

orchard is the standalone Python package for Orchard. Install it into a Python environment, call the embedded client directly from scripts or services, or start the optional OpenAI-compatible HTTP server when another process needs to talk to local models. It wraps the Proxy Inference Engine, a local C++ and Metal runtime built for streaming, continuous batching, multiple loaded models, structured output, tool calls, and multimodal inputs.

macOS 14+ | Apple Silicon | Python 3.12+ | Apache-2.0

Official docs | Quickstart | Client | Streaming | Responses | Server | Batching | Multimodal | Structured Output | Tool Use | Models

Install

uv venv
source .venv/bin/activate
uv pip install orchard

If you are not using uv, install inside a virtual environment:

python -m venv .venv
source .venv/bin/activate
pip install orchard

The first request downloads the Orchard engine binary and the model weights you ask for. The engine binary is cached under ~/.orchard/; Hugging Face model files use the normal Hugging Face cache.

Quickstart

Use the Python client directly when you are writing a Python app, notebook, worker, or evaluation job. You do not need to start the HTTP server for this path.

Create hello_orchard.py:

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-4-E2B-it"

with InferenceEngine(load_models=[MODEL]) as engine:
    client = engine.client()
    response = client.chat(
        MODEL,
        [{"role": "user", "content": "Write one sentence about local AI."}],
        temperature=0.0,
        max_generated_tokens=64,
    )
    print(response.text)

Run it:

python hello_orchard.py

For larger Macs, try google/gemma-4-E4B-it, meta-llama/Llama-3.1-8B-Instruct, or Qwen/Qwen3.5-4B.

Streaming

client.chat(..., stream=True) returns token deltas as the engine produces them.

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-4-E2B-it"

with InferenceEngine(load_models=[MODEL]) as engine:
    client = engine.client()
    stream = client.chat(
        MODEL,
        [{"role": "user", "content": "Count from one to five."}],
        stream=True,
        temperature=0.0,
        max_generated_tokens=64,
    )

    for delta in stream:
        if delta.content:
            print(delta.content, end="", flush=True)
    print()

Responses API

Use responses() when you want OpenAI Responses-style output objects, text deltas, reasoning items, and function-call items.

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-4-E2B-it"

with InferenceEngine(load_models=[MODEL]) as engine:
    client = engine.client()
    response = client.responses(
        MODEL,
        input="Explain why local inference is useful in two sentences.",
        temperature=0.0,
        max_output_tokens=96,
    )
    print(response.output_text)

For text-only streaming from Responses:

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-4-E2B-it"

with InferenceEngine(load_models=[MODEL]) as engine:
    client = engine.client()
    for chunk in client.responses_text(
        MODEL,
        input="Give me three concise debugging tips.",
        temperature=0.0,
        max_output_tokens=96,
    ):
        print(chunk, end="", flush=True)
    print()

Async

Every client path has an async form. Use achat(), aresponses(), and aresponses_text() inside async services.

import asyncio

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-4-E2B-it"


async def main() -> None:
    async with InferenceEngine() as engine:
        await engine.load_model(MODEL)
        client = engine.client()
        response = await client.achat(
            MODEL,
            [{"role": "user", "content": "Say hello from Orchard."}],
            temperature=0.0,
            max_generated_tokens=64,
        )
        print(response.text)


asyncio.run(main())

HTTP Server

Start the server only when another process, curl, or an OpenAI-compatible client needs to talk to Orchard over HTTP. The normal Python path is the client above.

orchard serve --model google/gemma-4-E2B-it

The default server listens on http://127.0.0.1:8000.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
response = client.chat.completions.create(
    model="google/gemma-4-E2B-it",
    messages=[{"role": "user", "content": "Hello from Orchard."}],
)
print(response.choices[0].message.content)

The server exposes:

Endpoint	Use
POST /v1/chat/completions	Chat Completions, streaming, batching, tools, structured output
POST /v1/responses	Responses objects, event streams, reasoning, tool calls, multimodal input
POST /v1/completions	Text completions
POST /v1/embeddings	Embeddings for supported models
GET /v1/models	Loaded model list
GET /health	Server health

Batching

Pass a list of conversations to schedule prompts together. Orchard returns one response per prompt in order.

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-4-E2B-it"

with InferenceEngine(load_models=[MODEL]) as engine:
    client = engine.client()
    responses = client.chat(
        MODEL,
        [
            [{"role": "user", "content": "Say hello politely."}],
            [{"role": "user", "content": "Give me a fun fact about space."}],
        ],
        temperature=0.0,
        max_generated_tokens=24,
    )

    for response in responses:
        print(response.text)

Sync, async, streaming, batching, and best-of-N are all supported. See orchard/clients/client.py.

For the HTTP API, send the same shape in messages:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemma-4-E2B-it",
    "messages": [
      [{"role": "user", "content": "Say hello politely."}],
      [{"role": "user", "content": "Give me a fun fact about space."}]
    ],
    "max_completion_tokens": 24,
    "temperature": 0.0
  }'

Multimodal

Use Responses-style content parts for images. Pass images as data URLs.

import base64
from pathlib import Path

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-3-4b-it"
IMAGE = Path("apple.jpg")

image_url = "data:image/jpeg;base64," + base64.b64encode(IMAGE.read_bytes()).decode()

with InferenceEngine(load_models=[MODEL]) as engine:
    client = engine.client()
    response = client.responses(
        MODEL,
        input=[
            {
                "type": "message",
                "role": "user",
                "content": [
                    {"type": "input_text", "text": "What is in this image?"},
                    {"type": "input_image", "image_url": image_url},
                ],
            }
        ],
        temperature=0.0,
        max_output_tokens=96,
    )
    print(response.output_text)

Structured Output

Use JSON Schema when the caller needs machine-readable output.

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-4-E2B-it"

schema = {
    "type": "object",
    "properties": {
        "capital": {"type": "string"},
        "population": {"type": "integer"},
    },
    "required": ["capital", "population"],
}

with InferenceEngine(load_models=[MODEL]) as engine:
    client = engine.client()
    response = client.responses(
        MODEL,
        input="What is the capital of France and its approximate population?",
        text={
            "format": {
                "type": "json_schema",
                "name": "city_info",
                "schema": schema,
                "strict": True,
            }
        },
        temperature=0.0,
        max_output_tokens=64,
    )
    print(response.output_text)

Tool Use

Tools use the Responses function schema. Non-streaming responses expose parsed function calls on response.tool_calls.

import json

from orchard.engine.inference_engine import InferenceEngine

MODEL = "google/gemma-4-E2B-it"

weather_tool = {
    "type": "function",
    "name": "get_weather",
    "description": "Get the current weather for a location.",
    "parameters": {
        "type": "object",
        "properties": {
            "location": {
                "type": "string",
                "description": "City name, e.g. San Francisco",
            }
        },
        "required": ["location"],
    },
}

with InferenceEngine(load_models=[MODEL]) as engine:
    client = engine.client()
    response = client.responses(
        MODEL,
        input="What is the weather in San Francisco?",
        tools=[weather_tool],
        tool_choice="required",
        temperature=0.0,
        max_output_tokens=128,
    )

    for call in response.tool_calls:
        print(call.name, json.loads(call.arguments))

Reasoning

Reasoning is model-dependent. For models with native thinking tokens, pass reasoning=True or an effort level.

response = client.responses(
    MODEL,
    input="Solve 23 * 47 and explain the steps briefly.",
    reasoning={"effort": "medium"},
    temperature=0.0,
    max_output_tokens=128,
)

Accepted effort values are minimal, low, medium, and high.

Production Use

Orchard is designed for production local services, not just one-off scripts. The same package covers notebooks, batch jobs, benchmark harnesses, and long-running agents that keep several models warm.

Capability	Path
Multiple loaded models	Start InferenceEngine(load_models=[...]) or orchard serve --model model-a model-b
Continuous batching	Send batched prompts or concurrent requests through the same engine process
Streaming	Use stream=True, responses_text(), or Server-Sent Events over HTTP
Structured output	Use response_format for Chat Completions or text.format for Responses
Tool use	Use tools, tool_choice, and max_tool_calls
Multimodal input	Use Responses content parts with input_text and input_image

The engine process is shared by Orchard clients on the machine. Stop it when you want a clean shutdown:

orchard engine stop

Update the engine binary:

orchard upgrade stable

Models

Orchard resolves local paths and Hugging Face repos on demand. The currently tested families include Gemma, Qwen, Llama, and Moondream.

Mac	Start with
8 GB unified memory	google/gemma-4-E2B-it
16-32 GB unified memory	google/gemma-4-E4B-it or Qwen/Qwen3.5-4B
32 GB+ unified memory	meta-llama/Llama-3.1-8B-Instruct

Other model families need a profile in Pantheon, which supplies chat templates, control tokens, and capability metadata.

Requirements

• macOS 14 or newer
• Apple Silicon Mac
• Python 3.12 or newer
• Disk space for the engine binary and model weights

Privacy

Inference runs locally on your Mac. Orchard downloads the engine binary and the model weights you request; prompts and outputs are not sent to a cloud inference API by Orchard.

Development

uv venv
source .venv/bin/activate
uv pip install -e ".[dev]"
pytest

For full engine/client verification inside the Proxy Company hyper-repo, run:

./scripts/pie_cycle.sh --py-only

Related

• Official Orchard docs
• orchard-rs for Rust apps that embed Orchard
• orchard-swift for Swift telemetry
• Pantheon
• Proxy Inference Engine

License

Apache-2.0