Siri AI is rolling out to macOS 27, even in the EU - despite this kind of Apple Intelligence not coming to European iPhones or iPads, at least not initially. Along with the Siri AI system, macOS exhibits direct access to AFM3 Core - the LLM at the foundation.
AFM3 command line client
macOS 27 comes with two primary ways to interact with this new type of Apple Intelligence: the new Siri app (more below) and the fm command line tool. This requires accepting the terms & conditions first:
% fm
YOU HAVE NOT AGREED TO THE APPLE FOUNDATION MODELS CLI LEGAL NOTICE & TERMS.
Agreeing to the Apple Foundation Models CLI Legal Notice & Terms applies to every user on the machine, so it must be run as a privileged user (e.g. ‘sudo fm license’).
“sudo fm license” references https://www.apple.com/legal/sla/ and lets users to accept these. However, “fm available” may still report model unavailability:
% fm available
System model unavailable: modelNotReady% fm chat
Error: The model is not available. Try again later.
This is markedly different from “deviceNotEligible”, so signals that the hardware is not the problem. After setting both the Siri language and the system language to English, it then seems necessary to enter the waitlist for Siri AI:
After that, AFM3 will start downloading invisibly in the background right away, even though the waitlist request is still pending:
(I was notified of Siri AI access through a push notification a couple of hours later
)
When the download is done (check the network load in the Activity Monitor), “fm available” will report:
% fm available
System model available
Using AFM3
Besides an interactive text-based chat interface (“fm chat”) and single shot prompts (“fm respond”), there is also a server (“fm serve”). This offers an OpenAI ChatCompletions compatible endpoint. I have extended my mlx_chat Gradio 6 chatbot UI to support this, including image support (AFM3 is a VLM).
AFM3 background
AFM3 has two distinct families: “Core” for on-device work, and “Cloud”. AFM 3 Core consists of “Core” proper - a 3 billion parameters dense model - and “Core Advanced”, a 20B sparse model, with 1 to 4 billion parameters active. As the article explains, the 20B model is not a standard Model-of-Experts (MoE) configuration, but routing decisions to experts are made on a per-prompt basis. The FM Teardown thus calls the “system” model shown by “fm available” (and the models ChatCompletion endpoint) a “router”.
Dynamic tracing on M3 showed that “fm serve” uses AFM3 Core Advanced - based on the “sparse” moniker in the filename.
Benchmark results
Apple itself doesn’t publish standard benchmark scores. I have thus run comparisons against Gemma 4 E4B myself (with the help of GPT-6 Astra, largely automated, so no guarantees) - see results in the appendix. Overall, Gemma E4B appears to be the stronger model, at the cost of (much) slower inference and higher memory use. Apple describes several use cases, while noting that the model “may not be suitable for handling all requests, like … create code”. But an MBPP+ score of 56.9% (versus 69.6% for Gemma 4 E4B) hardly means that it “can’t code” - users should simply not expect state-of-the-art coding performance.
Siri AI app
The Siri app on macOS 27 comes in two versions: the one with the colorful icon:
and the one with the grayscale icon - installed after having successfully cleared the waitlist and confusingly reporting as “Siri 1.0”:
This “Siri 1.0” looks like a barebones version of ChatGPT - but already supports image generation:
It also supports offline-requests to some degree - but will fall back to requiring an online connection. Interestingly, this app stopped working for me after a short while, and now reports “Siri AI is not available”, even though the “fm” CLI continues to work.
Appendix: AFM3 Core Advanced benchmark scores
TinyMMLU
AFM3 Core Advanced: 62%
Gemma 4 E4B: 69%
TinyGSM8K
AFM3 Core Advanced: 82%
Gemma 4 E4B: 94%
TinyHellaSwag
AFM3 Core Advanced: 66%
Gemma 4 E4B: 71%
TinyWinogrande
AFM3 Core Advanced: 55%
Gemma 4 E4B: 60%
ScienceQA 100-subset (multi-modal)
AFM3 Core Advanced: 79%
Gemma 4 E4B: 77%
MMStar 120-subset (multi-modal)
AFM3 Core Advanced: 48.3%
Gemma 4 E4B: 49.2%
MBPP+ (coding)
AFM3 Core Advanced: 56.9%
Gemma 4 E4B: 69.6%
TerminalBench 4.0 (agentic work)
AFM3 frequently got its context window exhausted and failed response formatting instruction-following
EuroEval subset - German
Grammar
AFM3 Core Advanced: 0.423
Gemma 4 E4B: 0.518
Sentiment
AFM3 Core Advanced: 0.539
Gemma 4 E4B: 0.578
Reading comprehension
AFM3 Core Advanced: 45.4%
Gemma 4 E4B: 48.9%