Add int8 (8-bit k-means-palettized) weight-only tier of both encoders

clap_audio_int8 / clap_text_int8: 8-bit per-tensor k-means palettization of the shipped fp16 weights (activations stay fp16), ~2x smaller. Chosen over linear int8 by measurement (text worst cosine vs fp16: kmeans 0.99873 vs linear 0.99571). int8-vs-fp16 embedding cosine worst: audio 0.99971, text 0.99873. Zero-shot ranking + cross-placement behavior unchanged. coremltools 9.0.

CHECKSUMS.sha256

@@ -14,3 +14,19 @@ edbf40e51518ad8f65a1d070f4c0e9b87fea644e6d69eb9eb55652e9fc697885  ./clap_text.ml
 6c68d30cf088f657428175055e25e01552c5af56769b119dc25b7463d7135ade  ./clap_text.mlpackage/Data/com.apple.CoreML/model.mlmodel
 7f4e15e9ccb0ffbc2341eec286e9d9934d3d3d8d6465dfddebed248bddc0e3dd  ./clap_text.mlpackage/Data/com.apple.CoreML/weights/weight.bin
 ff9a1b297cd1593982aff0eb36cf867f0cbb9af67eac749c0fe7545cd13450dc  ./clap_text.mlpackage/Manifest.json
+4455479a6d65004fe95d582246675cc9167c0b8fb6e0e673ad9db4009c0443de  ./clap_audio_int8.mlmodelc/analytics/coremldata.bin
+4bee9628bdf8821a391b32a7d23967c3c6712ae72088a318266583f35743ac33  ./clap_audio_int8.mlmodelc/coremldata.bin
+c5c466395cfb5a58ae9ed6f44e208492c59c6a1a245919928d08d87a8ffcf964  ./clap_audio_int8.mlmodelc/metadata.json
+1cdbbcc0911e9a9d427119e182dc3efa93d90b7159a614d372d397aff7861bb1  ./clap_audio_int8.mlmodelc/model.mil
+b3a37ec5550dcdd6932b314b830275ebcba013748421e1a517760b9afeabafb8  ./clap_audio_int8.mlmodelc/weights/weight.bin
+0dfa76a3a646e8001a923afb7e3c530791a7d2ec4840c4c0bcc1a4e0666d95e8  ./clap_audio_int8.mlpackage/Data/com.apple.CoreML/model.mlmodel
+b3a37ec5550dcdd6932b314b830275ebcba013748421e1a517760b9afeabafb8  ./clap_audio_int8.mlpackage/Data/com.apple.CoreML/weights/weight.bin
+25e72daa127949f1d7a09388d7498e987aa08a3a2dcb602555caf4fc6c462b96  ./clap_audio_int8.mlpackage/Manifest.json
+b914a93d50ae8336aad0977e28c4c1f84dccc39f7845fcd1660c43d565ca6fb7  ./clap_text_int8.mlmodelc/analytics/coremldata.bin
+95bc733c2b0d3d2fa64edf548f77a50f344c020cce382a0b96c46477ad4a0b84  ./clap_text_int8.mlmodelc/coremldata.bin
+0ed552101f045f5864c9d7212bffe09b478a2c1e8c39ec1d2732f95d206ee1df  ./clap_text_int8.mlmodelc/metadata.json
+0cc3ccdcf48e622a4701fa44e4c2096f0bead887461cb400e2b4af4e5641c2ee  ./clap_text_int8.mlmodelc/model.mil
+f181a595cefce402335499c32ea2f9727ef334afea9c592a2eabebb4172350a0  ./clap_text_int8.mlmodelc/weights/weight.bin
+468bdc0392942630945fc0e5ae65ba87f4576028d8899ba6f5125aeb122e69c5  ./clap_text_int8.mlpackage/Data/com.apple.CoreML/model.mlmodel
+f181a595cefce402335499c32ea2f9727ef334afea9c592a2eabebb4172350a0  ./clap_text_int8.mlpackage/Data/com.apple.CoreML/weights/weight.bin
+f6aed740178a9f3696c7dfb22fa9378b1ebfcbb3f167b5c2761ac43694779c42  ./clap_text_int8.mlpackage/Manifest.json

README.md

@@ -30,9 +30,13 @@ out of the graph entirely).
 |---|---|---|
 | `clap_audio.{mlpackage,mlmodelc}` | fp16 | HTSAT audio tower + `audio_projection` (spectrogram-input) |
 | `clap_text.{mlpackage,mlmodelc}` | fp16 | RoBERTa text tower + `text_projection` |
+| `clap_audio_int8.{mlpackage,mlmodelc}` | int8 | audio tower — 8-bit k-means-palettized weights (weight-only; activations fp16) |
+| `clap_text_int8.{mlpackage,mlmodelc}` | int8 | text tower — 8-bit k-means-palettized weights (weight-only; activations fp16) |
 
 `.mlpackage` is the canonical distributable; `.mlmodelc` is the compiled form the
-test suite loads directly. `CHECKSUMS.sha256` covers every file.
+test suite loads directly. `CHECKSUMS.sha256` covers every file. The `_int8`
+siblings are an **optional ~2× smaller** tier (see "int8 weight-only tier" below);
+the fp16 graphs remain the default.
 
 ### I/O contract (pinned from the artifact metadata)
 
@@ -88,6 +92,58 @@ time-major `[1001, 64]` → reshaped to `[1, 1, 1001, 64]`.
 
   fp16 is clean on every placement — **fp16 is shipped** (no fp32 fallback needed).
 
+## int8 weight-only tier (optional, ~2× smaller)
+
+Alongside the fp16 graphs this repo ships an **8-bit weight-only** tier of both
+towers — `clap_audio_int8` / `clap_text_int8`. Weights are stored 8-bit and
+dequantized to fp16 at runtime; **activations stay fp16** (identical graph maths,
+only the constant weights are compressed). Produced from the shipped fp16
+`.mlpackage`s by coremltools post-training compression — no reconversion, same
+source weights.
+
+- **Method: 8-bit k-means palettization, per-tensor**
+  (`coremltools.optimize.coreml.palettize_weights`, `nbits=8`, `mode="kmeans"`,
+  `weight_threshold=2048`). Chosen over linear-symmetric int8
+  (`linear_quantize_weights`) **by measurement**: on the text tower linear int8
+  costs up to **0.43%** cosine (worst 0.99571 vs fp16) while k-means costs
+  **0.13%** (worst 0.99873); on audio the two are close (0.99965 vs 0.99971).
+  k-means wins on both towers and is decisive on text, so it is the shipped `_int8`
+  tier. Tiny bias / LayerNorm-gain tensors (< 2048 elements) stay fp16; one RoBERTa
+  attention-mask fill constant (±inf) is likewise left uncompressed.
+- **int8-vs-fp16 embedding cosine** (identical inputs, CPU; the conversion
+  verification set — 10 real clips + 12 varied prompts including CJK — plus the
+  committed `golden_mel` fixture):
+
+  | tower | worst | mean |
+  |---|---|---|
+  | audio | 0.99970584 | 0.99982787 |
+  | text  | 0.99873250 | 0.99965990 |
+
+  Well inside a 0.5% cosine budget on both towers. (int8-vs-fp32-**source** worst:
+  audio 0.99967, text 0.99897 — barely further from the PyTorch source than fp16
+  itself, which is 0.99997 / 0.99988.)
+- **Zero-shot ranking unchanged.** The end-to-end gate (a ~192 s speech clip → 20
+  windows → aggregate → 4-anchor zero-shot score) returns the **identical ranking**
+  on int8 as on fp16: top label "This is a sound of a person speaking" at logit
+  **8.88** (fp16 8.82), same order for music / dog / rain, same margins.
+- **Placement unchanged.** Cross-compute-unit agreement holds (audio 0.99998, text
+  0.99994 — ≥ the 0.9999 band on All / CPU+NE / CPU+GPU / CPU). Audio still fails
+  `ANECCompile()` and falls back to GPU/CPU; text still compiles for the ANE — the
+  palettized weights change neither.
+- **Size** (`weight.bin` bytes, fp16 → int8):
+
+  | tower | fp16 | int8 | ratio |
+  |---|---|---|---|
+  | audio | 60,562,432 | 30,409,920 | 1.99× |
+  | text  | 251,665,792 | 125,746,688 | 2.00× |
+
+  Compiled `.mlmodelc`: audio 58 MB → 29 MB, text 240 MB → 120 MB.
+
+Same toolchain pin as the fp16 conversion (coremltools 9.0; scikit-learn provides
+the k-means step). The int8 tier is a pure recompression of the pinned fp16
+artifacts, so its provenance and licensing are exactly those of the fp16 graphs
+below.
+
 ## Placement guidance (measured, never marketed)
 
 - **text**: compiles for and runs on the ANE/GPU/CPU; fp16-clean on all.

clap_audio_int8.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4455479a6d65004fe95d582246675cc9167c0b8fb6e0e673ad9db4009c0443de
+size 243

clap_audio_int8.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4bee9628bdf8821a391b32a7d23967c3c6712ae72088a318266583f35743ac33
+size 613

clap_audio_int8.mlmodelc/metadata.json

@@ -0,0 +1,80 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "CLAP audio encoder (HTSAT + audio_projection), spectrogram-input — int8 weight-only (int8_kmeans)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "audio_embeds",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float16, Palettized (8 bits))",
+    "modelParameters" : [
+
+    ],
+    "author" : "clapkit T1: laion\/clap-htsat-unfused audio tower (HTSAT + audio_projection), pre-norm | int8 weight-only (int8_kmeans) tier",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.mul" : 12,
+      "Ios17.linear" : 77,
+      "Ios17.transpose" : 79,
+      "Ios16.constexprLutToDense" : 93,
+      "Ios17.matmul" : 25,
+      "Ios17.conv" : 1,
+      "Ios17.gather" : 20,
+      "Ios17.add" : 41,
+      "Ios17.sliceByIndex" : 18,
+      "Ios17.batchNorm" : 1,
+      "Ios17.concat" : 4,
+      "Ios16.softmax" : 12,
+      "Ios16.gelu" : 12,
+      "Ios16.reduceMean" : 1,
+      "Ios16.relu" : 1,
+      "Ios17.reshape" : 130,
+      "Ios17.layerNorm" : 29,
+      "Ios17.cast" : 2
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 1 × 1001 × 64)",
+        "shortDescription" : "",
+        "shape" : "[1, 1, 1001, 64]",
+        "name" : "input_features",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-18",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "clap_audio_int8_kmeans",
+    "method" : "predict"
+  }
+]

clap_audio_int8.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3a37ec5550dcdd6932b314b830275ebcba013748421e1a517760b9afeabafb8
+size 30409920

clap_audio_int8.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0dfa76a3a646e8001a923afb7e3c530791a7d2ec4840c4c0bcc1a4e0666d95e8
+size 239348

clap_audio_int8.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b3a37ec5550dcdd6932b314b830275ebcba013748421e1a517760b9afeabafb8
+size 30409920

clap_audio_int8.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "4DD6422A-5E33-4740-A410-AD78A02EA811": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        },
+        "9B4A7D10-2F98-4170-9A92-0E30990F6E18": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        }
+    },
+    "rootModelIdentifier": "4DD6422A-5E33-4740-A410-AD78A02EA811"
+}

clap_text_int8.mlmodelc/analytics/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b914a93d50ae8336aad0977e28c4c1f84dccc39f7845fcd1660c43d565ca6fb7
+size 243

clap_text_int8.mlmodelc/coremldata.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95bc733c2b0d3d2fa64edf548f77a50f344c020cce382a0b96c46477ad4a0b84
+size 617

clap_text_int8.mlmodelc/metadata.json

@@ -0,0 +1,94 @@
+[
+  {
+    "metadataOutputVersion" : "3.0",
+    "shortDescription" : "CLAP text encoder (RoBERTa + text_projection) — int8 weight-only (int8_kmeans)",
+    "outputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Float32",
+        "formattedType" : "MultiArray (Float32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "text_embeds",
+        "type" : "MultiArray"
+      }
+    ],
+    "storagePrecision" : "Mixed (Float16, Palettized (8 bits))",
+    "modelParameters" : [
+
+    ],
+    "author" : "clapkit T1: laion\/clap-htsat-unfused text tower (RoBERTa + text_projection), pre-norm | int8 weight-only (int8_kmeans) tier",
+    "specificationVersion" : 8,
+    "mlProgramOperationTypeHistogram" : {
+      "Ios17.notEqual" : 1,
+      "Ios17.reshape" : 48,
+      "Ios17.logicalAnd" : 1,
+      "Ios16.softmax" : 12,
+      "Ios17.matmul" : 24,
+      "Ios17.transpose" : 48,
+      "Select" : 3,
+      "Ios17.add" : 41,
+      "Ios17.sliceByIndex" : 1,
+      "Ios17.gatherAlongAxis" : 1,
+      "Ios17.gatherNd" : 1,
+      "Ios17.gather" : 3,
+      "Ios17.layerNorm" : 25,
+      "Ios16.gelu" : 12,
+      "Ios17.cast" : 6,
+      "Ios16.constexprLutToDense" : 79,
+      "Ios17.tanh" : 1,
+      "Ios16.relu" : 1,
+      "Ios17.linear" : 75,
+      "Ios17.greaterEqual" : 2,
+      "Ios16.cumsum" : 1,
+      "Ios17.mul" : 13
+    },
+    "computePrecision" : "Mixed (Float16, Float32, Int32, Int8, UInt16)",
+    "isUpdatable" : "0",
+    "stateSchema" : [
+
+    ],
+    "availability" : {
+      "macOS" : "14.0",
+      "tvOS" : "17.0",
+      "visionOS" : "1.0",
+      "watchOS" : "10.0",
+      "iOS" : "17.0",
+      "macCatalyst" : "17.0"
+    },
+    "modelType" : {
+      "name" : "MLModelType_mlProgram"
+    },
+    "inputSchema" : [
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "input_ids",
+        "type" : "MultiArray"
+      },
+      {
+        "hasShapeFlexibility" : "0",
+        "isOptional" : "0",
+        "dataType" : "Int32",
+        "formattedType" : "MultiArray (Int32 1 × 512)",
+        "shortDescription" : "",
+        "shape" : "[1, 512]",
+        "name" : "attention_mask",
+        "type" : "MultiArray"
+      }
+    ],
+    "userDefinedMetadata" : {
+      "com.github.apple.coremltools.conversion_date" : "2026-07-18",
+      "com.github.apple.coremltools.source" : "torch==2.5.1",
+      "com.github.apple.coremltools.version" : "9.0",
+      "com.github.apple.coremltools.source_dialect" : "TorchScript"
+    },
+    "generatedClassName" : "clap_text_int8_kmeans",
+    "method" : "predict"
+  }
+]

clap_text_int8.mlmodelc/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f181a595cefce402335499c32ea2f9727ef334afea9c592a2eabebb4172350a0
+size 125746688

clap_text_int8.mlpackage/Data/com.apple.CoreML/model.mlmodel

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:468bdc0392942630945fc0e5ae65ba87f4576028d8899ba6f5125aeb122e69c5
+size 160269

clap_text_int8.mlpackage/Data/com.apple.CoreML/weights/weight.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f181a595cefce402335499c32ea2f9727ef334afea9c592a2eabebb4172350a0
+size 125746688

clap_text_int8.mlpackage/Manifest.json

@@ -0,0 +1,18 @@
+{
+    "fileFormatVersion": "1.0.0",
+    "itemInfoEntries": {
+        "77FCF9B8-1057-4833-877D-1D93B4F0B351": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Weights",
+            "name": "weights",
+            "path": "com.apple.CoreML/weights"
+        },
+        "C5028B06-BB64-4D12-B086-117B2D0852DB": {
+            "author": "com.apple.CoreML",
+            "description": "CoreML Model Specification",
+            "name": "model.mlmodel",
+            "path": "com.apple.CoreML/model.mlmodel"
+        }
+    },
+    "rootModelIdentifier": "C5028B06-BB64-4D12-B086-117B2D0852DB"
+}