# Core AI model zoo

> Community ports of open models to Apple's Core AI runtime (`.aimodel`, iOS/macOS 27), each
> with the recipe that produced it, plus a knowledge base of verified findings about the
> runtime itself. Apple documents the API surface; these notes cover what it does when you run
> it — thresholds, failure modes, and measured numbers — and are specific about what was
> verified, on what hardware, and how.

## Start here

- [AGENTS.md](https://john-rocky.github.io/coreai-model-zoo/AGENTS.html): the porting contract in one file — why conversion is not format
  conversion, the two gates every port gets, the traps agents specifically hit, and which
  actions stay a human's call.
- [README.md](https://john-rocky.github.io/coreai-model-zoo/): the catalog itself — every model, its card, its Hugging Face
  repo, and the one-line Swift call that runs it.
- [models/index.json](https://john-rocky.github.io/coreai-model-zoo/models/index.json): the same catalog machine-readable. Per recipe:
  `status` (is the configuration recorded), `source_model` (what it was converted from), and
  `gate_transcript` (is the numerical check against the original published, and where).
- [PORTING.md](https://john-rocky.github.io/coreai-model-zoo/PORTING.html): the full walk from a Hugging Face checkpoint to a verified
  bundle on an iPhone, with two worked examples.
- [SECURITY.md](https://john-rocky.github.io/coreai-model-zoo/SECURITY.html): what the integrity story is, including the parts that are
  absent — pinned revisions rather than signatures, and no checksum manifest.
- [Source repository](https://github.com/john-rocky/coreai-model-zoo): the conversion scripts, the gates, and the recipes behind
  every page here. The site renders the same files; nothing is written for it separately.
- [The Art of Core AI](https://john-rocky.github.io/the-art-of-core-ai/): the long-form version
  — a free book built from these same measurements, for reading start to finish rather than
  looking one thing up.

## Orientation

- [undocumented-answers.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/undocumented-answers.html): the questions Apple's docs leave open, with measured answers: the AOT threshold, the iOS-only dynamic-KV miscompile at seq ≥2048, whether a 4B fits on the ANE, what the chunk threshold really dials, and why a hand-written kernel forces a single-token export
- [coreai-overview.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/coreai-overview.html): what Core AI is, the 3 Apple repos, the .aimodel format, the PyTorch → .aimodel → Swift-runtime pipeline
- [conversion-guide.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/conversion-guide.html): converting a PyTorch model to .aimodel: the canonical TorchConverter API + the gotchas that cost real time
- [ship-playbook.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/ship-playbook.html): the end-to-end runbook: converted .aimodel → CoreAIKit Swift engine → app → on-device (AOT + sideload + headless self-test for RTF) → publish (HF + zoo + post). The stage checklist + cross-cutting traps (gate-before-port, JIT→AOT, tokenizer retag), validated shipping Parakeet in one session
- [compute-units-and-authoring.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/compute-units-and-authoring.html): ANE vs GPU vs CPU: the static/BC1S/Conv2d/per-head/fp16 ANE rules vs the dynamic/fused/custom-kernel GPU rules, the macOS↔iOS export split, and the PSNR verification gates. Read this to choose a target
- [performance-ceiling.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/performance-ceiling.html): reality check: where Core AI LLM decode tops out (Mac GPU near its ceiling, MLX gap structural, fusion closed), what AOT does/doesn't do, and why ANE is an energy play not a speed one. Read before chasing a "dramatic speed" win

## GPU-now track (speed)

- [pipelined-engine.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/pipelined-engine.html): read this first for decode speed: riding Apple's coreai-pipelined engine = 3.5× over a hand-rolled per-token loop with ZERO custom kernels (qwen3.5: Mac 204 tok/s, iPhone 50.3–51.5). The decode-only loop-free export, the extra-states engine patch, the chunk=1 / warmup-256 traps, LUT-vs-linear int8, oracle gating, and what fits/doesn't (Gemma 4's PLE doesn't — yet)
- [int8-head-and-decode-measurement.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/int8-head-and-decode-measurement.html): the int8 LM head decode lever (untie + absmax int8, ~half the per-token read) and how to measure the win honestly: controlled-bench vs in-app, VLM image_embeds dilution (+48% text core → +36% VLM), thermal-robust ratios, and the SwiftUI O(n²) re-decode that masquerades as a slow model
- [fm-provider.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/fm-provider.html): zoo models behind Apple's LanguageModelSession (WWDC 339): CoreAILanguageModel(resourcesAt:) = the whole integration (verified, incl. hybrid/SSM bundles on the patched pipelined engine), plus the own-conformance recipe that adds tool calling (verified round trip) and the protocol traps (dead prewarm, no guided generation on pipelined, re-prefill tax)
- [custom-metal-kernels.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/custom-metal-kernels.html): TorchMetalKernel (WWDC 325): the API, the register-then-add order, MSL embedded in the .aimodel, GPU-only, what to (and not to) kernelize. Still the tool when a model CAN'T ride the pipelined engine (e.g. Gemma 4)
- [rwkv7-recurrent-linear-attention-coreai.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/rwkv7-recurrent-linear-attention-coreai.html): the no-KV playbook for pure-recurrent / linear-attention LLMs (RWKV-7, Mamba-2, gated-delta-net): the matrix-state decode recurrence lowers to standard ops (no custom kernel); O(1) fixed-size states, no KV cache wired via SSMState + fused end-of-step writes; the pipelined engine can't drive a no-KV model (positional keyCache/valueCache) → a custom backend binding states BY NAME (like BitVLA); the recurrence-protecting quant recipe (int8keepproj); teacher-forced gating. Shipped RWKV-7 1.5B, iPhone 17 Pro 25.2 tok/s
- [tensorops-quantized-kernels.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/tensorops-quantized-kernels.html): the layer UNDER custom kernels (WWDC 330): TensorOps quantized matmul (int4/int8 = OS 26; fp4/fp8/int2 + E8M0 scale planes = OS 27), cooperative tensors, the FlashAttention recipe, and the M5/A19 GPU neural accelerator — the compute-bound/prefill lever hand-rolled MSL can't reach
- [prefix-cache-kv-reuse.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/prefix-cache-kv-reuse.html): cross-turn KV reuse: turn-2 TTFT 0.23 s vs 23.3 s = 101× at a 4k context, proven lossless (greedy token-identical). The orthogonal speed lever nobody's shipping on-device yet
- [spec-decode-design.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/spec-decode-design.html): speculative decoding on the pipelined decode bundles: the only lever past the decode bandwidth wall (verify K tokens per forward); design + feasibility. GDN-hybrid (Qwen3.5/3.6) static-S verify companion: [spec-decode-hybrid-verify-design.md](spec-decode-hybrid-verify-design.md)
- [tensorops-zoo-impact-and-kernel-wins.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/tensorops-zoo-impact-and-kernel-wins.html): applied survey: where TensorOps quantized kernels and pure custom-Metal wins actually pay across the zoo

## Benchmarks & comparisons

- [apple-models-bench.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/apple-models-bench.html): measured numbers for Apple's own coreai-models export recipes — the README Apple didn't write (21 recipes, zero official numbers)
- [coreai-vs-mlx-speed.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/coreai-vs-mlx-speed.html): every measured Core AI–vs–MLX decode comparison (same M4 Max, same protocol) + the causal decomposition of the gap: where Core AI wins, where it structurally can't, and why

## ANE-later track (when the beta KV-write bug lifts + int4 head + AOT)

- [aot-and-specialization.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/aot-and-specialization.html): specialization, AIModelCache / AIModel.specialize(), and AOT compile (xcrun coreai-build compile → .aimodelc, --preferred-compute neural-engine). The first-run-latency mitigation path
- [compression-reference.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/compression-reference.html): coreai-opt quantization & palettization API reference (int4/int8, granularity, mixed-precision, joint); the LM-head/embedding lever
- [coreai-beta-mpsgraph-kvwrite-bug.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/coreai-beta-mpsgraph-kvwrite-bug.html): the data-indexed in-graph KV write SIGSEGV (FB23024751 / apple#5): platform-agnostic (GPU too), host-cache workaround

## Agent & app layer (FM framework / App Intents / Evaluations / security)

- [spotlight-rag-third-party.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/spotlight-rag-third-party.html): running Apple's WWDC26 SpotlightSearchTool (local RAG as one Tool) behind a third-party zoo model via KitLanguageModel: only .toolCalling is needed (not guided generation), the tool returns metadata-not-body (hydrate with a companion fetch_note tool), guidance level is a token gate, and the thinking-model /no_think mitigation. Verified example: coreai-kit/Examples/SpotlightChat
- [dynamic-profiles-local-models.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/dynamic-profiles-local-models.html): WWDC26 DynamicProfile (242) routing between two local zoo models (0.6B triage ↔ 4B expert) in one LanguageModelSession, fully on-device/airplane-mode — the config Apple's on-device↔PCC demo doesn't show. The body-purity rule, switch re-prefill cost, two-resident-model footprint, and why the model-decision channel must be guided-gen (not a tool) on the stock engine. Example: agent-demos/DualProfileChat
- [visual-intelligence-third-party-model.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/visual-intelligence-third-party-model.html): running YOUR own model (CLIP / RF-DETR) behind the system Visual Intelligence camera/screenshot search (WWDC26 297): IntentValueQuery + SemanticContentDescriptor, model-agnostic by construction (no model param, no capability, no entitlement), and the real gate — running a model in the query's background-launch memory budget. Example: coreai-kit/Examples/VisualIntel
- [agentic-security-checklist.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/agentic-security-checklist.html): pre-ship checklist for on-device LLM agent apps (WWDC 347+343): indirect prompt injection, the Lethal Trifecta, .onToolCall/.historyTransform guardrails, App-Intents risk-based confirmation + authenticationPolicy + OwnershipProvidingEntity
- [evaluations-framework.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/evaluations-framework.html): Apple's Evaluations framework (WWDC 298/299/335) mapped to this project's oracle/margin gates; the disallowed-trajectory injection test; a Vault-style on-device eval suite

## Runtime & decode internals

- [compression.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/compression.html): this project's LLM-specific empirical compression notes (int8 floor, per-subsystem sensitivity); pairs with compression-reference.md
- [stateful-kv-cache.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/stateful-kv-cache.html): stateful decode export, dual/hybrid KV state, the sliding-window ring buffer, the dynamic prefill+decode graph
- [swift-runtime.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/swift-runtime.html): the Core AI Swift API, driving .aimodel from Swift, non-standard architectures, macOS/Xcode 27 setup (incl. running Xcode 27 beta without sudo)

## Model port notes (per-architecture lessons)

- [bitcpm-ternary-1.58bit.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/bitcpm-ternary-1.58bit.html): 1.58-bit ternary MiniCPM4-8B: the zoo's first sub-int8 packed-GEMM Metal kernel; an 8B running in ~2.1 GB on the iPhone GPU
- [bitvla-1.58bit-vla.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/bitvla-1.58bit-vla.html): 1.58-bit Vision-Language-Action (robotics): image + instruction → 7-DoF actions, fully on-device
- [minicpm5-1b.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/minicpm5-1b.html): the clean-LlamaForCausalLM recipe done end-to-end (hybrid Think/No-Think, untied head, 128K) — the most reusable conversion template in the zoo
- [youtu-mla-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/youtu-mla-port.html): dense DeepSeek-style MLA at 2B on iPhone: latent-KV attention with an absorbed flash-decode kernel
- [diffusion-llms-dllm.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/diffusion-llms-dllm.html): masked-diffusion LLMs (LLaDA): parallel canvas denoising, bidirectional attention, no KV cache — and how that maps to Core AI graphs
- [gliner2-pii.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/gliner2-pii.html): DeBERTa-v3 (disentangled attention) NER / schema-driven zero-shot extraction; the on-device PII-redaction model. [unlimited-ocr-rswa-static-decode.md](unlimited-ocr-rswa-static-decode.md) — the document-OCR trio: a Glm4v variant, whole-page parsing on stock Qwen2-VL, and an R-SWA MoE on the stock runtime with static-shape stateful decode
- [video-world-models-vjepa2.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/video-world-models-vjepa2.html): V-JEPA 2: a self-supervised video world model as on-device action classification (16-frame clips)
- [gemma4-mixedbit-qat-transplant.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/gemma4-mixedbit-qat-transplant.html): extracting Google's mobile mixed-bit QAT weights and transplanting them into Core AI bundles
- [gemma4-ple-static-input-fm-stack.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/gemma4-ple-static-input-fm-stack.html): Gemma 4's per-layer-embedding table as a static graph input, loaded behind FoundationModels
- [timesfm-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/timesfm-port.html): TimesFM 2.5: the zoo's first time-series forecasting foundation model (stateless graph + host RevIN DSP)
- [esam3-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/esam3-port.html): EfficientSAM3: a dropped port (device-verified but redundant vs the official SAM 3) — kept for what transferred

## Image generation & editing (diffusion)

- [zimage-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/zimage-port.html): Z-Image-Turbo, a 6B Single-Stream DiT text-to-image — and why it ships Mac-only
- [glm-image-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/glm-image-port.html): GLM-Image: the zoo's first AR + diffusion hybrid (9B GLM-4 AR writes visual prior tokens → 7B flow-matching DiT renders)
- [flux2-in-context-editing.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/flux2-in-context-editing.html): FLUX.2 [klein] instruction editing + multi-reference composition with no separate editing model and no ControlNet

## Audio (the new modality — model-specific port notes)

- [qwen2.5-omni-audio-understanding.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/qwen2.5-omni-audio-understanding.html): Qwen2.5-Omni's Thinker as on-device audio understanding (describes sounds, *not* ASR): de-dynamizing the Whisper-style encoder (ragged cu_seqlens chunk attention → one batched fixed attention, bit-exact), TMRoPE collapsing to 1-D, a bit-exact vDSP mel (DFT-as-matmul), audio embeds on a static buffer, and the two payoffs — AOT clean-mmap weights dodge the iOS jetsam dirty limit (4.5 GB decoder, 5930 MB free), and a fixed-shape encoder runs on the ANE (0.99 cos → byte-identical text) where the dynamic decoder can't
- [whisper-asr-fixed-decode.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/whisper-asr-fixed-decode.html): Whisper large-v3-turbo as on-device ASR (the zoo's first official speech-to-text): why the official recipe's single-step [1,1] decoder graph can't transcribe, why a dynamic-length re-export recompiles every step (~15 s/token), and the fix — a fixed 128-token decoder window (pad, read logits at the real last position; causal attention ignores the padding; constant shape compiles once → 0.18 s/token, token-exact). Plus the Swift log-mel frontend (n_fft 400 = DFT-as-matmul) and single-main→GPU routing
- [kokoro-tts.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/kokoro-tts.html): Kokoro-82M (StyleTTS2 + iSTFTNet), the zoo's first text-to-speech: 3 bundles cut at the one data-dependent length + host DSP, the weight_norm-random-init bug (non-determinism → suspect weight loading), variable length via masked-unrolled bi-LSTM + frame-masked InstanceNorm + host STFT, the Core AI op rewrites (ConvTranspose1d/iSTFT → zero-insert + conv1d), and why an unrolled-LSTM model runs faster on the CPU than the GPU
- [adcsr-super-resolution.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/adcsr-super-resolution.html): AdcSR, the zoo's first super-resolution (one-step diffusion-GAN, pruned SD-2.1 + half VAE decoder): why fp16 NaNs (SD-2.1 attention overflow, upcast_attention ignored by the SDPA processor) so fp32 ships, the global (not per-tile) color-match, host-side tiling with an input-size cap, and the CoreGraphics bytesPerRow/no-y-flip traps
- [sam3-promptable-segmentation.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/sam3-promptable-segmentation.html): SAM 3 as on-device text-prompt segmentation (the zoo's first official segmenter): the segmenter *bundle* format + one-call official CoreAIImageSegmenter runtime, float16≈float32 fidelity (Δ≤1e-4) verified with the image-segmenter CLI, redistributing a gated model under the SAM License (ship the LICENSE), and the platform box-origin trap when compositing masks
- [depth-anything-3-monocular-depth.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/depth-anything-3-monocular-depth.html): Depth Anything 3, the zoo's first depth model (DINOv2 ViT + DPT head, single static square graph): why the input is raw [0,1] (ImageNet norm folded in-graph) and the double-normalization trap that faked a cos-0.9 "engine bug" / "noise" / "non-square bug" for a day; engine cos 1.000000 at any shape; the squish + resize-back contract and why r≈0.98 vs official is faithful (= the model's own 504-vs-518 variance); the RoPE-const / cache-free / pos-embed-dtype export patches; fp16 via .half() not autocast. Sample: [scripts/depth_anything_3_sample.py](scripts/depth_anything_3_sample.py)
- [voxcpm-tts.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/voxcpm-tts.html): VoxCPM-0.5B, the zoo's first diffusion TTS: a family of graphs (LM + diffusion + VAE + vocoder) resolved as one catalog model
- [chatterbox-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/chatterbox-port.html): Chatterbox: the zoo's first zero-shot voice-cloning TTS and first multi-network port
- [music-generation-stable-audio.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/music-generation-stable-audio.html): Stable Audio Open: text→music latent diffusion (T5 cond + DiT + VAE) on-device
- [sortformer-speaker-diarization.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/sortformer-speaker-diarization.html): streaming speaker diarization: export only the neural core, port the streaming loop + AOSC speaker cache to the host
- [vibevoice-multispeaker-tts.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/vibevoice-multispeaker-tts.html): VibeVoice, the zoo's first multi-speaker / dialogue TTS: dual-LM next-token diffusion, why the speech feedback loop forces fp16, and why the expectFrequentReshapes hint must be off on iOS
- [melband-source-separation.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/melband-source-separation.html): Mel-Band RoFormer, the zoo's first source separation: complex → real arithmetic, and folding STFT/iSTFT into the graph as constant DFT matmuls so the on-device host needs no FFT

## Everything else in the knowledge base

- [accel-levers-survey-and-plan.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/accel-levers-survey-and-plan.html): Accel levers — industry survey + zoo maximal-optimization plan (2026-07-01)
- [coreai-torch-041-ir-incident.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/coreai-torch-041-ir-incident.html): The coreai-torch 0.4.0 IR-location incident, and how to recover / re-verify
- [cross-runtime-quality-benchmarking.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/cross-runtime-quality-benchmarking.html): Cross-runtime quality benchmarking: how to not measure your own harness
- [dense-int4km-flagship-session-findings.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/dense-int4km-flagship-session-findings.html): Dense-path int4km coverage + flagship Qwen3.6-35B — session findings & methods (2026-07-01)
- [flagship-full-tuning-stack.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/flagship-full-tuning-stack.html): Flagship dramatic-speedup plan — the full stackable tuning set (2026-07-01)
- [gemma4-litertlm-to-official-migration.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/gemma4-litertlm-to-official-migration.html): Gemma-4 E2B mixed-bit QAT — migration from .litertlm extraction to the official HF checkpoint
- [gemma4-raw-metal-a19-levers.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/gemma4-raw-metal-a19-levers.html): Gemma4 raw-Metal loop on A19 — findings + the remaining (parked) levers
- [gemma4-raw-metal-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/gemma4-raw-metal-port.html): Gemma-4-E2B raw-Metal port — hand-written mixed-bit decode loop (ship doc)
- [gemma4-wna8o8-requires-int8-activations.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/gemma4-wna8o8-requires-int8-activations.html): Gemma-4 E2B mobile QAT (wNa8o8): the weights require their int8 activation path
- [glm-ocr-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/glm-ocr-port.html): GLM-OCR → Core AI: port notes
- [kernel-campaign-handoff-2026-07-02.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/kernel-campaign-handoff-2026-07-02.html): Kernel campaign — session handoff (2026-07-02)
- [lfm2audio-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/lfm2audio-port.html): LFM2.5-Audio — unified speech↔speech (ASR + TTS) on Core AI (port notes)
- [mineru-port.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/mineru-port.html): MinerU2.5-Pro → Core AI — port notes
- [nanbeige4.2-coreai-support.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/nanbeige4.2-coreai-support.html): Nanbeige4.2-3B on Core AI
- [quant-d-fp4-qat-feasibility-and-plan.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/quant-d-fp4-qat-feasibility-and-plan.html): Stream D — Quantization (FP4 + QAT-int4): feasibility, first result, and plan
- [raw-metal-loop-playbook.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/raw-metal-loop-playbook.html): Raw-Metal loop playbook — when leaving the engine pays, and when it doesn't
- [spec-decode-c-feasibility-and-plan.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/spec-decode-c-feasibility-and-plan.html): Stream C — Speculative decoding: feasibility verdict + implementation plan (2026-07-01)
- [spec-decode-c1-handoff-2026-07-03.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/spec-decode-c1-handoff-2026-07-03.html): Spec-decode C1 — DONE (2026-07-03): the loop now beats shipped 15.9 tok/s
- [spec-decode-hybrid-verify-design.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/spec-decode-hybrid-verify-design.html): Spec-decode on GDN hybrids (Qwen3.5/3.6 family) — static-S verify design (2026-07-02)
- [ssm-hybrid-decode-int4-prefill-findings.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/ssm-hybrid-decode-int4-prefill-findings.html): SSM-hybrid decode, int4, and chunked prefill on Core AI — session findings (2026-07-10/11)
- [ternary-chunked-prefill.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/ternary-chunked-prefill.html): Tiled ternary GEMM + chunked prefill — lifting the S=1 tax off custom-kernel bundles
- [unlimited-ocr-rswa-static-decode.md](https://john-rocky.github.io/coreai-model-zoo/knowledge/unlimited-ocr-rswa-static-decode.html): Static-shape stateful decode + stock-runtime VLM (Unlimited-OCR)

## Optional

- [CONTRIBUTING.md](https://john-rocky.github.io/coreai-model-zoo/CONTRIBUTING.html): what an accepted port must clear, and the device
  gate — the one step a contributor without an iOS 27 device can hand back.
- [BENCHMARKS.md](https://john-rocky.github.io/coreai-model-zoo/BENCHMARKS.html): community-submitted device measurements, explicitly not
  a controlled-environment benchmark.