fm — Apple Foundation Models CLI: Reverse-Engineering Notes (macOS 27)

Field notes on /usr/bin/fm, the Apple Foundation Models command-line tool that shipped with macOS 27, and the on-device + Private Cloud Compute model stack behind it.

Everything here was derived from publicly-shipped macOS binaries using standard tools (strings, otool, codesign, curl) plus live runs of the fm CLI itself. No proprietary weights, source, or assets are redistributed. See DISCLAIMER.

Snapshot from an early macOS 27 build (internal SDK macosx27.0.internal, FoundationModels.framework v2.0.51). Identifiers and behavior will change in later builds.

---

TL;DR

• fm is a thin Apple-signed CLI over the public FoundationModels.framework. It adds no models — it's a developer front door to the same stack Siri and Apple Intelligence use.
• It exposes two execution tiers, not two models: system (on-device) and pcc (Private Cloud Compute).
• Behind those tiers, the ModelCatalog enumerates a real model zoo: on-device language tiers instruct_9m / 85m / 300m / 3b, a separate code-model family, vision and speech models, and a 193-use-case cloud/PCC adapter catalog.
• The 3B on-device tier ships as a dense + sparse pair — cross-referenced to Apple's post, that's AFM 3 Core (3B dense) + AFM 3 Core Advanced (20B sparse, 1–4B active) — with .draft speculative-decoding twins for low latency (~0.26 s short completions measured).
• fm does not route between models — it pins one tier. The real local↔cloud routing is a Siri-side orchestration layer (ORCHNLRouterBridge → serverFallback → PCC), gated by attestation (TC2XPCTrustedRequestProtocol, coherence tokens).
• fm serve is a working OpenAI Chat-Completions API server (streaming + non-streaming, real token accounting, TCP or Unix socket) — drop-in for OpenAI-compatible clients, but chat-only: any request with a tools array is 400-rejected (Invalid tool definition), so agent frameworks must be configured to omit tool definitions entirely.

---

Documents

Doc	What's in it
00 — Field Report	Consolidated, evidence-graded conclusions; corrections to common assumptions; the big-picture diagram; honest limits.
01 — CLI Reference	Every command and flag, the in-REPL slash commands, and the full fm serve API with verified request/response examples.
02 — Model Architecture	The ModelCatalog layering, on-device size tiers + adapters, the code/vision/speech families, the PCC tier, trust/attestation, and how Siri routes local↔cloud.
03 — Official Cross-Reference	Our findings mapped to Apple's third-gen Foundation Models post: model-name reconciliation (Core / Core Advanced / Cloud / Cloud Pro / ADM), what Apple confirms, what we add, and corrections it forces.
Original Notes	The earlier behavioral session notes that seeded this (host specs redacted), kept for provenance.

Evidence grading

Findings are tagged so you can tell measured facts from inference:

• [OBSERVED] — produced by running commands on a real macOS 27 host.
• [STRINGS] — recovered from symbol/string analysis of the signed Apple binaries (high confidence the capability exists in the OS; not necessarily reachable from fm).
• [INFERRED] — reasoned conclusion, flagged as such.

---

Raw evidence (artifacts/)

File	Contents
fm_model_ids.txt	1,037 com.apple.fm.* model-catalog identifiers (the core evidence)
pcc_use_cases.txt	193 cloud/PCC instruct_server use-case adapters
cloud_endpoints.txt	afm.aiml.apple.com / smoot.apple.com inference endpoints
siri_pcc_models.txt	llm_siri_pcc + lw_planner_v1..v5
movie_schema.json	Sample guided-generation schema (round-tripped live)
logs/serve.log	Captured fm serve output (loopback only)

---

Reproduce it yourself

On a macOS 27 machine:

# CLI surface
fm --help
for c in respond chat token-count schema serve available quota-usage; do fm $c --help; done

# Live on-device generation
fm respond --greedy --no-stream 'Say hello in three words.'
fm available
fm quota-usage

# Guided / constrained generation
fm schema object --name Movie --string title --integer year --boolean rewatchable > movie.json
fm respond --no-stream --schema movie.json 'Give me a classic sci-fi film.'

# OpenAI-compatible server
fm serve --port 1976 &
curl -s localhost:1976/v1/models
curl -s localhost:1976/v1/chat/completions -H 'Content-Type: application/json' \
  -d '{"model":"system","messages":[{"role":"user","content":"PONG?"}],"stream":false}'

Binary / catalog inspection (the model-id list came from this):

codesign -dvvv /usr/bin/fm
otool -L /usr/bin/fm
CACHE=/System/Volumes/Preboot/Cryptexes/OS/System/Library/dyld
for f in "$CACHE"/dyld_shared_cache_arm64e*; do strings -a "$f"; done \
  | grep -aoE 'com\.apple\.fm\.[a-z0-9._]+' | sort -u

The actual model weight files under /System/Library/AssetsV2/com_apple_MobileAsset_UAF_FM_* are SIP-protected (restricted flag) and not readable, so size labels are Apple's own catalog identifiers, not measured parameter counts.

---

Contributing

These are notes from one macOS 27 build — findings from other builds, hardware, or a PCC-enabled host are very welcome. See CONTRIBUTING for what's useful and the ground rules (no proprietary content, keep the evidence grading, show your build).

What's not here

• No model weights, no extracted Apple source, no proprietary assets.
• No live PCC traces — PCC was attestation-gated on the test host, so the cloud-tier details are from string analysis, not cloud calls.

See DISCLAIMER for scope, copyright, and trademark notes.