Apple Foundation Models SDK for Python: Local Apple Intelligence Inside the Hermes Agent

In a post on Habr, Sber engineers presented a native provider that connects Apple Intelligence's local model to the autonomous Python agent Hermes through the Apple Foundation Models SDK for Python. By their measurements, the average response time on short queries was 2.8 seconds, and on a 50-question benchmark the adapter produced 46 correct answers.

What is the Foundation Models SDK for Python

The Apple Foundation Models SDK for Python is a library that gives Python code direct access to Apple Intelligence's local model. Apple released it in February 2026: the model loads into the process, the KV-cache stays in RAM, and inference runs on Apple Silicon chips without any calls to the cloud.

The library quickly found an audience among developers. According to GitHub, it already has 1.2k stars, and Apple presented the latest version, 0.2.0, with image support at the WWDC26 conference.

• Library released in February 2026
• 1.2k stars on GitHub
• Version 0.2.0 with image support unveiled at WWDC26
• Fully local inference on Apple Silicon
• KV-cache stored in process memory

Why an adapter for Hermes was needed

Until now, there was no ready-made native provider for Python agents. Existing wrappers around Foundation Models emulate the OpenAI interface, but they do not let the local model act as the agent's primary engine with all of its machinery: the rate limiter, fallback to other models, and auditing. The Swift agent iClaw already used Foundation Models directly, but that option remained closed to more general-purpose Python agents.

The Hermes adapter closes that gap: Apple's local model becomes the agent's standard inference backend rather than an external service behind an HTTP wrapper. That means requests to it pass through the same limit, fallback, and logging logic as calls to cloud models.

What the benchmarks showed

On a test set of 50 questions, the provider scored 46 correct answers. The questions came from five well-known datasets — MMLU, BoolQ, GSM8K, BIG-Bench, and HellaSwag — so the test covered knowledge, reasoning, and arithmetic at once. Average response time on short queries was 2.8 seconds, which is noticeably fast for fully local inference on a laptop.

The combination of speed and accuracy is the case's main selling point: the agent answers with no network latency and without spending API tokens.

"Why not try adding a

Foundation Models SDK for Python adapter to Hermes and get an average response time of 2.8 seconds on short queries and 46 correct answers on a 50-question benchmark?" — from the Sber team's article on Habr.

What this means

Local models have come close to serving as the primary engine for autonomous agents: no cloud, no API fees, and acceptable latency. For scenarios where privacy and offline operation matter, the combination of Apple Silicon and Python is shifting from an exotic setup to a working tool.

Frequently asked questions

What is the Apple Foundation Models SDK for Python?

It is a Python library released by Apple in February 2026 that gives direct access to Apple Intelligence's local model: the model loads into the process and runs on Apple Silicon without the cloud. It has 1.2k stars on GitHub, and the current version is 0.2.0 with image support.

What results did the Hermes adapter show?

On a 50-question benchmark (MMLU, BoolQ, GSM8K, BIG-Bench, HellaSwag), the provider produced 46 correct answers, and the average response time on short queries was 2.8 seconds with fully local inference.