orchard 2026.5.4

Python client for Orchard, a compute platform for Apple Silicon

Orchard

100% local, OpenAI-compatible LLM inference for Apple Silicon. Multi-model serving, prefix caching, continuous batching. No cloud APIs, no data leaves your machine.

macOS 14+ · Apple Silicon (M1+) · Python 3.12+ · Apache-2.0

Features

• Drop-in OpenAI API — /v1/chat/completions, /v1/responses, /v1/embeddings, /v1/models
• Fast — C++ inference engine with prefix caching and continuous batching
• Multi-model — load Qwen, Llama, and Gemma side-by-side; swap between them per request
• Multimodal — vision, tool calling, thinking; native where the model was trained, grammar-constrained where it wasn't
• OpenAI Responses API — streaming events for reasoning, tool calls, and messages
• Use from anything — curl, Python (openai SDK), Rust (orchard-rs), or any OpenAI-compatible client

Getting started

pip install orchard
orchard serve --model google/gemma-4-E4B-it

First run downloads the PIE engine binary (~2 GB) and the model weights from HuggingFace. Subsequent runs start in seconds.

Then point anything at http://localhost:8000/v1:

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "google/gemma-4-E4B-it",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Or with the OpenAI Python SDK:

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-E4B-it",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)

Supported models

Orchard downloads open-weights models from HuggingFace on demand. Any model whose family has a profile in Pantheon works out of the box.

Model	Size (BF16)	Modalities	Best for
google/gemma-4-E4B-it (default)	~8 GB	text, vision	Multimodal, native thinking, tool calls
Qwen/Qwen3.5-4B	~9 GB	text	256k context, native thinking, tool calls
meta-llama/Llama-3.1-8B-Instruct	~16 GB	text	General-purpose, trained tool calls
google/gemma-3-4b-it	~8 GB	text, vision	Multimodal chat
google/gemma-4-E2B-it	~5 GB	text, vision	Fits on 8 GB Macs
moondream/moondream3-preview	~9 GB	vision	Pointing, detection, captioning

By hardware

Your Mac	Recommended
M1 / M2 / M3 (8 GB)	google/gemma-4-E2B-it
M-Pro / Max (16–32 GB)	google/gemma-4-E4B-it, Qwen/Qwen3.5-4B
M-Max / Ultra (32+ GB)	meta-llama/Llama-3.1-8B-Instruct + a small model hot-loaded

Need quantized weights? Pass any mlx-community/... repo directly. Orchard resolves any HuggingFace repo whose architecture belongs to a supported family.

OpenAI Responses API

Orchard implements the OpenAI Responses API — the successor to Chat Completions. You get structured streaming events for reasoning, tool calls, and messages, plus per-item state and lifecycle events.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="not-needed")
stream = client.responses.create(
    model="google/gemma-4-E4B-it",
    input="Explain quantum tunneling in three sentences.",
    reasoning={"effort": "medium"},
    stream=True,
)
for event in stream:
    print(event)

Supports response.output_text.delta, response.reasoning.delta, response.function_call_arguments.delta, structured output via JSON Schema, and max_tool_calls for bounded tool loops.

Python SDK (no HTTP)

For embedded use — skip the HTTP layer:

from orchard.engine.inference_engine import InferenceEngine

async with InferenceEngine() as engine:
    client = engine.client()
    response = await client.achat(
        "google/gemma-4-E4B-it",
        messages=[{"role": "user", "content": "Hello!"}],
    )
    print(response.text)

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

Privacy

Orchard runs entirely on your Mac. No telemetry, no analytics, no phone-home.

Surface	Status	What it does
Inference	✅ Local	All generation on-device via PIE (C++, Metal)
Chat templates	✅ Local	Rendered from Pantheon profiles bundled in the package
Model weights	✅ One-time	HuggingFace Hub → ~/.cache/huggingface/
Engine binary	✅ One-time	GitHub release → ~/.orchard/bin/
Telemetry	✅ None	No tracking SDKs — verify with grep -r analytics orchard/

How it works

Orchard is the Python layer over a stack built for Apple Silicon:

• PIE — C++ inference engine: prefix caching, continuous batching, multi-model scheduling
• PAL — Metal GPU kernels
• PSE — grammar-constrained generation for tool calls, structured output, and thinking
• Pantheon — chat templates and capability manifests, shared across all Orchard SDKs

The Python package handles IPC, model resolution, HuggingFace downloads, prompt rendering, and the FastAPI server.

CLI

orchard serve --model <hf-repo> [--host 127.0.0.1] [--port 8000]
orchard serve --models model-a model-b model-c    # preload multiple
orchard upgrade [stable|nightly]                  # update engine binary
orchard engine stop                               # kill background engine

Requirements

• macOS 14+, Apple Silicon (M1 or newer)
• Python 3.12+
• ~2 GB free disk for the engine binary, plus model weights

Related

• orchard-rs — Rust client
• orchard-swift — Swift telemetry client
• Pantheon — model profiles
• PIE — the engine

License

Apache-2.0