Apple Foundation Models SDK for Python: локальная Apple Intelligence в агенте Hermes

Engineers from Sber showed off, in a blog post on Habr, a native provider that connects the local Apple Intelligence model to the autonomous Python agent Hermes via the Apple Foundation Models SDK for Python. According to their measurements, the average response time for short queries was 2.8 seconds, and on a benchmark of 50 questions the adapter delivered 46 correct answers.

What is Foundation Models SDK for Python

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

The library quickly gained a following among developers. According to GitHub, it already has 1.2 thousand stars, and Apple unveiled the latest version, 0.2.0, with image support, at the WWDC26 conference.

• Library release — February 2026
• 1.2 thousand stars on GitHub
• Version 0.2.0 with image support — at WWDC26
• Inference is fully local, on Apple Silicon
• The KV cache is stored in process memory

Why Hermes needed an adapter

No ready-made native provider for Python agents existed before this. Wrappers around Foundation Models emulate the OpenAI interface, but they don't let the local model work as the agent's primary engine — with all its mechanisms: rate limiting, fallback to other models, and auditing. Direct use of Foundation Models already existed in the Swift agent iClaw, but for more general-purpose Python agents this capability had remained closed off.

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

What results did the benchmarks show

On a test set of 50 questions, the provider scored 46 correct answers. The questions were drawn from five well-known benchmark sets — MMLU, BoolQ, GSM8K, BIG-Bench, and HellaSwag — meaning they tested knowledge, logic, and arithmetic all at once. The average reaction time for short queries was 2.8 seconds — for fully local inference on a laptop, that's noticeably fast.

The combination of speed and accuracy is the main argument for the case: the agent responds without network latency and without spending tokens on API calls.

"Why not try adding an adapter based on the

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

What this means

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

Frequently Asked Questions

What is Apple Foundation Models SDK for Python?

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

What results did the Hermes adapter show?

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