Performance guide
Decode optimization boundaries
Which checkpoints qualify for MTP and EAGLE3, what silently falls back, and which automatic MLX changes ship in the build.
Engineering notes
The working record behind AFM.
Design constraints, measured performance work, migration maps, and negative results. These are repository documents—not manufactured blog posts.
Read guide →
Investigation
Why exact KV replay is unsafe
Boundary-sensitive cache behavior, the failed replay variants, and the current safe policy.
Architecture note
Apple-native endpoint conventions
The error, availability, CORS, model-field, and discovery contract shared by Vision, Speech, and embeddings.
Feasibility study
Paged attention and radix caching
Why true page-backed KV storage is an architectural change, not a label for chunked snapshots.
Performance record
DeepSeek V4: rejected experiments
A maintained negative-results ledger with exact-output controls, paired runs, and promotion thresholds.
Migration map
AFMKit and WWDC 26 Foundation Models
Where the portable API already has a seam for newer Apple capabilities and what remains intentionally unimplemented.