@qvac/sdk 0.20.1 → 0.21.0

package/NOTICE

@@ -71,6 +71,8 @@ Third-Party Model Licenses
     https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
   bci-embedder
     https://github.com/tetherto/qvac/releases/download/bci-test-assets-v0.1.0/bci-embedder.bin
+  coremldata
+    https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
   cosyvoice3-campplus-f32
     https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   cosyvoice3-flow-bf16
@@ -136,7 +138,7 @@ Third-Party Model Licenses
   LightOnOCR-2-1B-ocr-soup-GGUF
     https://huggingface.co/noctrex/LightOnOCR-2-1B-ocr-soup-GGUF
   Llama_3.2_1B_Intruct_Tool_Calling_V2-GGUF
-    https://huggingface.co/mav23/Llama_3.2_1B_Intruct_Tool_Calling_V2-GGUF
+    https://huggingface.co/mradermacher/Llama_3.2_1B_Intruct_Tool_Calling_V2-GGUF
   MedPsy-1.7B-GGUF
     https://huggingface.co/qvac/MedPsy-1.7B-GGUF
   MedPsy-4B-GGUF
@@ -145,6 +147,14 @@ Third-Party Model Licenses
     https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   mmproj-unlimited-ocr-Q8_0
     https://huggingface.co/baidu/Unlimited-OCR
+  model
+    https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
+  moss-codec-decoder-f16
+    https://huggingface.co/OpenMOSS-Team/MOSS-TTS
+  moss-codec-encoder-f16
+    https://huggingface.co/OpenMOSS-Team/MOSS-TTS
+  moss-tts-delay-f16
+    https://huggingface.co/OpenMOSS-Team/MOSS-TTS
   nb-tiny-ggml-model
     https://huggingface.co/NbAiLab/nb-whisper-tiny
   parler-indic-f16
@@ -211,6 +221,8 @@ Third-Party Model Licenses
     https://huggingface.co/HuggingFaceVLA/smolvla_libero
   SmolVLM2-500M-Video-Instruct-GGUF
     https://huggingface.co/ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+  Ternary-Bonsai-2-27B-gguf
+    https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf
   TranslatePsy-AfriSLM-0.8B-Q4-GGUF
     https://huggingface.co/qvac/TranslatePsy-AfriSLM-0.8B-Q4-GGUF
   TranslatePsy-AfriSLM-0.8B-Q8-GGUF
@@ -251,6 +263,8 @@ Third-Party Model Licenses
     https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged
   Wan2.2-TI2V-5B-Turbo-GGUF
     https://huggingface.co/hum-ma/Wan2.2-TI2V-5B-Turbo-GGUF
+  weight
+    https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
 
 --- bsd-3-clause (BSD 3-Clause License) ---
 
@@ -277,30 +291,58 @@ Third-Party Model Licenses
 
   AfriqueGemma-4B-GGUF
     https://huggingface.co/mradermacher/AfriqueGemma-4B-GGUF
+  coremldata
+    https://huggingface.co/nvidia/parakeet-ctc-0.6b
+  coremldata
+    https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
+  coremldata
+    https://huggingface.co/nvidia/parakeet-tdt-1.1b
+  coremldata
+    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
+  model
+    https://huggingface.co/nvidia/parakeet-ctc-0.6b
+  model
+    https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
+  model
+    https://huggingface.co/nvidia/parakeet-tdt-1.1b
+  model
+    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
+  parakeet_realtime_eou_120m-v1.f16
+    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
+  parakeet_realtime_eou_120m-v1.q4_0
+    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
+  parakeet_realtime_eou_120m-v1.q8_0
+    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
   parakeet-ctc-0.6b.f16
     https://huggingface.co/nvidia/parakeet-ctc-0.6b
   parakeet-ctc-0.6b.q4_0
     https://huggingface.co/nvidia/parakeet-ctc-0.6b
   parakeet-ctc-0.6b.q8_0
     https://huggingface.co/nvidia/parakeet-ctc-0.6b
-  parakeet-eou-120m-v1.f16
-    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
-  parakeet-eou-120m-v1.q4_0
-    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
-  parakeet-eou-120m-v1.q8_0
-    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
   parakeet-tdt-0.6b-v3.f16
     https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
   parakeet-tdt-0.6b-v3.q4_0
     https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
   parakeet-tdt-0.6b-v3.q8_0
     https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
+  parakeet-tdt-1.1b.f16
+    https://huggingface.co/nvidia/parakeet-tdt-1.1b
+  parakeet-tdt-1.1b.q8_0
+    https://huggingface.co/nvidia/parakeet-tdt-1.1b
   sortformer-4spk-v1.f16
     https://huggingface.co/nvidia/diar_sortformer_4spk-v1
   sortformer-4spk-v1.q4_0
     https://huggingface.co/nvidia/diar_sortformer_4spk-v1
   sortformer-4spk-v1.q8_0
     https://huggingface.co/nvidia/diar_sortformer_4spk-v1
+  weight
+    https://huggingface.co/nvidia/parakeet-ctc-0.6b
+  weight
+    https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3
+  weight
+    https://huggingface.co/nvidia/parakeet-tdt-1.1b
+  weight
+    https://huggingface.co/nvidia/parakeet_realtime_eou_120m-v1
 
 --- gemma (Gemma Terms of Use) ---
 
@@ -362,6 +404,8 @@ Third-Party Model Licenses
     https://huggingface.co/ResembleAI/chatterbox
   chatterbox-t3-turbo
     https://huggingface.co/ResembleAI/chatterbox-turbo
+  coremldata
+    https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
   de-base-ggml-model-f16
     https://huggingface.co/aware-ai/whisper-base-german
   DeepSeek-V4-Flash-0731-GGUF
@@ -384,16 +428,18 @@ Third-Party Model Licenses
     https://huggingface.co/ChristianAzinn/gte-large-gguf
   gte-large-gguf
     https://huggingface.co/ChristianAzinn/gte-large-gguf
-  indic-conformer-ctc.f16
+  indic-conformer-600m-multilingual.f16
     https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
-  indic-conformer-ctc.q4_0
+  indic-conformer-600m-multilingual.q4_0
     https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
-  indic-conformer-ctc.q8_0
+  indic-conformer-600m-multilingual.q8_0
     https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
   lavasr-denoiser-f16
     https://huggingface.co/spaces/YatharthS
   lavasr-enhancer-f16
     https://huggingface.co/spaces/YatharthS
+  model
+    https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
   pt-tiny-ggml-model-f16
     https://huggingface.co/DavyCosta701/whisper-tiny-pt-bpra
   taew2_2_f16
@@ -404,6 +450,8 @@ Third-Party Model Licenses
     https://voiceinput.futo.org/VoiceInput/tiny_en_acft_q8_0.bin
   vae-BF16
     https://huggingface.co/ACE-Step/Ace-Step1.5
+  weight
+    https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
   whisper-vad
     https://huggingface.co/ggml-org/whisper-vad
   whisper.cpp
@@ -618,6 +666,12 @@ Third-Party Model Licenses
 
 --- nvidia-open-model-license ---
 
+  coremldata
+    https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2.1
+  coremldata
+    https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
+  coremldata
+    https://huggingface.co/nvidia/parakeet-unified-en-0.6b
   diar_streaming_sortformer_4spk-v2.1.f16
     https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2.1
   diar_streaming_sortformer_4spk-v2.1.q4_0
@@ -628,6 +682,12 @@ Third-Party Model Licenses
     https://huggingface.co/nvidia/GR00T-N1.7-LIBERO
   groot-q8_vf16
     https://huggingface.co/nvidia/GR00T-N1.7-3B
+  model
+    https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2.1
+  model
+    https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
+  model
+    https://huggingface.co/nvidia/parakeet-unified-en-0.6b
   nemotron-3.5-asr-streaming-0.6b.f16
     https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
   nemotron-3.5-asr-streaming-0.6b.q4_0
@@ -640,6 +700,12 @@ Third-Party Model Licenses
     https://huggingface.co/nvidia/parakeet-unified-en-0.6b
   parakeet-unified-en-0.6b.q8_0
     https://huggingface.co/nvidia/parakeet-unified-en-0.6b
+  weight
+    https://huggingface.co/nvidia/diar_streaming_sortformer_4spk-v2.1
+  weight
+    https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
+  weight
+    https://huggingface.co/nvidia/parakeet-unified-en-0.6b
 
 --- openrail ---
 
@@ -682,7 +748,6 @@ the underlying models.
 
 Source URLs for original weights are recorded in each model's
 metadata entry within the registry.
-
 =========================================================================
 JavaScript Dependencies
 =========================================================================
@@ -722,8 +787,6 @@ JavaScript Dependencies
     https://github.com/tetherto/qvac
   @qvac/logging@0.1.1
     https://github.com/tetherto/qvac
-  @qvac/model-fit@0.12.2
-    https://github.com/tetherto/qvac
   @qvac/ocr-ggml@0.24.3
     https://github.com/tetherto/qvac
   @qvac/rag@0.8.1

package/README.md

@@ -22,6 +22,13 @@ For AI/LLM tools, use [https://docs.qvac.tether.io/llms-full.txt](https://docs.q
 
 See https://docs.qvac.tether.io/sdk/getting-started/installation
 
+On Node.js, the SDK installs and launches its own Bare worker through `bare-runtime`.
+Its dependency range must satisfy `@qvac/inference`'s `engines.bare` requirement.
+When upgrading the SDK, reinstall dependencies with the existing lockfile so the
+package manager can replace an older, incompatible runtime. Remove any override
+or resolution that forces the worker below that requirement; installing a newer
+global `bare` executable does not change the SDK's worker dependency.
+
 ## Quickstart
 
 1. Create the examples workspace:
@@ -143,16 +150,16 @@ initialized, the event omits the resource block.
 ## Pre-download model fit assessment
 
 Use `assessModelFit` to check, before downloading anything, whether models are
-likely to fit in this device's memory. It reads generated catalog metadata plus a
-fresh memory sample, and for a single candidate the registry's weightless
-description of the artifact, so the engine's own fitter can answer — no weights,
-no load:
+likely to fit in this device's memory. Each candidate is described as `loadModel`
+takes it. It reads generated catalog metadata plus a fresh memory sample, and for
+a single candidate the registry's weightless description of the artifact, so the
+engine's own fitter can answer — no weights, no load:
 
 ```ts
 import { assessModelFit, QWEN3_8B_INST_Q4_K_M } from '@qvac/sdk'
 
 const result = await assessModelFit({
-  models: [{ model: QWEN3_8B_INST_Q4_K_M, workload: { kind: 'llm', contextTokens: 8192 } }],
+  models: [{ modelSrc: QWEN3_8B_INST_Q4_K_M, modelType: 'llm', modelConfig: { ctx_size: 8192 } }],
   execution: 'sequential',
   policy: 'interactive-v1'
 })
@@ -187,6 +194,33 @@ console.log(stats?.audioDuration, stats?.realTimeFactor)
 `session.stats` resolves to `undefined` when the engine does not report
 statistics.
 
+Batch `transcribe()` calls also expose terminal statistics:
+
+```ts
+const call = transcribe({ modelId, audioChunk })
+const text = await call
+const stats = await call.stats
+console.log(text, stats?.encoderUsedCoreml)
+```
+
+`call.stats` resolves to `undefined` when the engine reports no statistics and
+rejects if transcription fails.
+
+## Parakeet Core ML encoders on Apple devices
+
+Load a supported Parakeet registry constant as usual. On macOS and iOS, the SDK
+downloads any complete, published Core ML encoder bundle with its GGUF and
+places the `.mlmodelc` directory where the native addon discovers it. This is
+available for TDT 0.6B v3, Unified English 0.6B, EOU 120M v1, and streaming
+Sortformer v2.1. The download includes the sidecar weights, so first load takes
+more time and disk space. Other platforms download only the GGUF. If the
+sidecar is unavailable or cannot load, inference uses the GGUF encoder.
+Sortformer v2.1 also needs a GGUF with the
+`parakeet.model_variant=sortformer-streaming-v2.1-aosc` metadata. Caching both
+sidecars alone does not activate Core ML. See
+[`examples/asr/parakeet-unified-coreml.ts`](examples/asr/parakeet-unified-coreml.ts)
+for a local check of a Unified GGUF and its encoder bundle.
+
 ## Examples
 
 In the `./examples` subdirectory, you will find scripts demonstrating how to use all SDK functionalities. To try any of them:

package/dist/examples/advisory-model-fit.d.ts

@@ -1,15 +1,16 @@
 /**
- * Advisory llama.cpp fit check (QVAC-22629).
+ * Advisory model fit check.
  *
- * Before a completion or embedding load, the SDK runs `@qvac/model-fit` in one
- * disposable Bare child and projects whether the exact configuration it is
- * about to load will fit in device memory.
+ * Before a load, the SDK runs the fitter belonging to the engine that would
+ * run it and projects whether the exact configuration it is about to load will
+ * fit in device memory. Desktop runs it in a disposable Bare child; mobile,
+ * which cannot spawn one, runs it on a worker thread.
  *
  * The result is ADVISORY. It never blocks a load. `does-not-fit` is logged and
  * the ordinary load path runs unchanged. Crashes, timeouts, malformed
  * responses, unsupported configurations, and internal errors all resolve to
- * "no evidence" and are equally non-blocking. Nothing consumes the verdict
- * yet — this PR only produces it.
+ * "no evidence" and are equally non-blocking. No verdict changes the load;
+ * `getLoadedModelInfo` returns it as `fitProbe`.
  *
  * The verdict is emitted on the SDK server log stream, not to stdout, so this
  * example subscribes to `loggingStream({ id: SDK_LOG_ID })` and reprints the
@@ -19,9 +20,8 @@
  * What fits on this machine (Apple M4 Pro, 24 GiB unified memory)
  * ---------------------------------------------------------------------------
  *
- * Measured with `@qvac/model-fit@0.8.0` (the first release carrying the
- * qvac-fabric#214 memory-reporting fix) at the default 1024 MiB margin. The
- * fitter budgets against what the machine can actually keep resident
+ * Measured at the default 1024 MiB margin. The fitter budgets against what the
+ * machine can actually keep resident
  * (total − wired − compressor: 17.4 GiB on this machine at idle), not the raw
  * RAM figure.
  *
@@ -30,10 +30,12 @@
  *     gpt-oss-20B   Q4_K_M  10.8 GiB  @  32k ctx
  *     gte-large     fp16     0.6 GiB  (embedding, context pinned to 512)
  *
- *   PROJECTED NOT TO FIT — try any of these to see a `does-not-fit` verdict
+ *   PROJECTED NOT TO FIT — with `gpu_layers: 99` pinned
  *     gpt-oss-20B   Q4_K_M  10.8 GiB  @ 128k ctx with an f32 KV cache ← below
- *     Gemma 4 31B   Q4_K_M  18.3 GiB  @   1k ctx   ← the weights alone do not
- *                                                    fit with gpu_layers: 99
+ *
+ * These rows were measured with `gpu_layers` pinned at 99. Loads no longer pin
+ * the layer count, and the `does-not-fit` row has not been re-measured
+ * unpinned.
  *
  * gpt-oss-20B at 128k context FITS with the default KV cache and DOES NOT FIT
  * once `cache-type-k`/`cache-type-v` are set to `f32`. Same model, same
@@ -42,10 +44,9 @@
  *
  * Two boundaries worth understanding when reading verdicts:
  *
- * 1. The verdict answers for a PLACEMENT, not a model. The SDK's schema
- *    default is `gpu_layers: 99` ("everything on the GPU"), so `does-not-fit`
- *    means "not at this placement". Omit `gpu_layers` and the fitter is free
- *    to move layers to the CPU side instead.
+ * 1. The verdict answers for a PLACEMENT, not a model. With `gpu_layers`
+ *    unset, the fitter may move layers to the CPU side; setting it pins the
+ *    layer count, and `does-not-fit` then means "not at this placement".
  * 2. A `does-not-fit` configuration can still RUN on macOS when the OS
  *    compresses and pages hard enough — and a `fits` configuration right at
  *    the boundary can still fail at first decode under memory pressure.

package/dist/examples/advisory-model-fit.js

@@ -1,15 +1,16 @@
 /**
- * Advisory llama.cpp fit check (QVAC-22629).
+ * Advisory model fit check.
  *
- * Before a completion or embedding load, the SDK runs `@qvac/model-fit` in one
- * disposable Bare child and projects whether the exact configuration it is
- * about to load will fit in device memory.
+ * Before a load, the SDK runs the fitter belonging to the engine that would
+ * run it and projects whether the exact configuration it is about to load will
+ * fit in device memory. Desktop runs it in a disposable Bare child; mobile,
+ * which cannot spawn one, runs it on a worker thread.
  *
  * The result is ADVISORY. It never blocks a load. `does-not-fit` is logged and
  * the ordinary load path runs unchanged. Crashes, timeouts, malformed
  * responses, unsupported configurations, and internal errors all resolve to
- * "no evidence" and are equally non-blocking. Nothing consumes the verdict
- * yet — this PR only produces it.
+ * "no evidence" and are equally non-blocking. No verdict changes the load;
+ * `getLoadedModelInfo` returns it as `fitProbe`.
  *
  * The verdict is emitted on the SDK server log stream, not to stdout, so this
  * example subscribes to `loggingStream({ id: SDK_LOG_ID })` and reprints the
@@ -19,9 +20,8 @@
  * What fits on this machine (Apple M4 Pro, 24 GiB unified memory)
  * ---------------------------------------------------------------------------
  *
- * Measured with `@qvac/model-fit@0.8.0` (the first release carrying the
- * qvac-fabric#214 memory-reporting fix) at the default 1024 MiB margin. The
- * fitter budgets against what the machine can actually keep resident
+ * Measured at the default 1024 MiB margin. The fitter budgets against what the
+ * machine can actually keep resident
  * (total − wired − compressor: 17.4 GiB on this machine at idle), not the raw
  * RAM figure.
  *
@@ -30,10 +30,12 @@
  *     gpt-oss-20B   Q4_K_M  10.8 GiB  @  32k ctx
  *     gte-large     fp16     0.6 GiB  (embedding, context pinned to 512)
  *
- *   PROJECTED NOT TO FIT — try any of these to see a `does-not-fit` verdict
+ *   PROJECTED NOT TO FIT — with `gpu_layers: 99` pinned
  *     gpt-oss-20B   Q4_K_M  10.8 GiB  @ 128k ctx with an f32 KV cache ← below
- *     Gemma 4 31B   Q4_K_M  18.3 GiB  @   1k ctx   ← the weights alone do not
- *                                                    fit with gpu_layers: 99
+ *
+ * These rows were measured with `gpu_layers` pinned at 99. Loads no longer pin
+ * the layer count, and the `does-not-fit` row has not been re-measured
+ * unpinned.
  *
  * gpt-oss-20B at 128k context FITS with the default KV cache and DOES NOT FIT
  * once `cache-type-k`/`cache-type-v` are set to `f32`. Same model, same
@@ -42,10 +44,9 @@
  *
  * Two boundaries worth understanding when reading verdicts:
  *
- * 1. The verdict answers for a PLACEMENT, not a model. The SDK's schema
- *    default is `gpu_layers: 99` ("everything on the GPU"), so `does-not-fit`
- *    means "not at this placement". Omit `gpu_layers` and the fitter is free
- *    to move layers to the CPU side instead.
+ * 1. The verdict answers for a PLACEMENT, not a model. With `gpu_layers`
+ *    unset, the fitter may move layers to the CPU side; setting it pins the
+ *    layer count, and `does-not-fit` then means "not at this placement".
  * 2. A `does-not-fit` configuration can still RUN on macOS when the OS
  *    compresses and pages hard enough — and a `fits` configuration right at
  *    the boundary can still fail at first decode under memory pressure.
@@ -98,8 +99,7 @@ try {
     // Unloaded before the next phase, so the second verdict is measured on an
     // idle machine and stays comparable to the fixture tables above. Leaving it
     // loaded would shift the verdict: the check reserves resident weight bytes
-    // through the fit margin, and since model-fit 0.8.0 the fit child also sees
-    // system-wide wired memory.
+    // through the fit margin, and the fitter also sees system-wide wired memory.
     await unloadModel({ modelId: smallModelId, clearStorage: false });
     watchVerdicts();
     // 2. A load the fitter projects NOT to fit. The point of this example is that

package/dist/examples/asr/parakeet-indic-conformer-filesystem.js

@@ -4,14 +4,14 @@
  * Usage:
  *   bun run examples/asr/parakeet-indic-conformer-filesystem.ts <wav-file> <language> [gguf]
  *
- * Loads a single GGUF checkpoint (`PARAKEET_INDIC_CONFORMER_CTC_Q8_0` by
+ * Loads a single GGUF checkpoint (`PARAKEET_INDIC_CONFORMER_600M_Q8_0` by
  * default) and transcribes with the batch `transcribe` API. `language` is
  * required (e.g. `hi`, `ta`) because Indic Conformer CTC masks the vocab
  * with `parakeet.ctc.lang_*` ranges. English Parakeet CTC ignores this field.
  *
  * Audio should be 16 kHz mono PCM in a WAV container.
  */
-import { loadModel, unloadModel, transcribe, PARAKEET_INDIC_CONFORMER_CTC_Q8_0 } from '@qvac/sdk';
+import { loadModel, unloadModel, transcribe, PARAKEET_INDIC_CONFORMER_600M_Q8_0 } from '@qvac/sdk';
 const args = process.argv.slice(2);
 if (!args[0] || !args[1]) {
     console.error('Usage: bun run examples/asr/parakeet-indic-conformer-filesystem.ts ' +
@@ -22,7 +22,7 @@ if (!args[0] || !args[1]) {
 }
 const audioFilePath = args[0];
 const language = args[1];
-const parakeetModelSrc = args[2] ?? PARAKEET_INDIC_CONFORMER_CTC_Q8_0;
+const parakeetModelSrc = args[2] ?? PARAKEET_INDIC_CONFORMER_600M_Q8_0;
 try {
     console.log('▸ Loading Indic Conformer CTC model...');
     const modelId = await loadModel({

package/dist/examples/asr/parakeet-unified-coreml.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=parakeet-unified-coreml.d.ts.map

package/dist/examples/asr/parakeet-unified-coreml.js

@@ -0,0 +1,113 @@
+/**
+ * Transcribe a WAV file with Parakeet Unified and its Core ML encoder sidecar.
+ *
+ * From packages/sdk:
+ *   bun run examples/asr/parakeet-unified-coreml.ts \
+ *     /path/to/parakeet-unified-en-0.6b.q4_0.gguf [16-kHz-wav-file]
+ *
+ * The GGUF must be a regenerated Unified artifact. The five compiled Core ML
+ * files are downloaded through the SDK registry and staged beside the GGUF.
+ * The default audio is examples/audio/sample-16khz.wav.
+ */
+import { downloadAsset, getModelInfo, loadModel, transcribe, unloadModel } from '@qvac/sdk';
+import { link, mkdir, mkdtemp, open, rm, symlink } from 'node:fs/promises';
+import { tmpdir } from 'node:os';
+import { basename, join, resolve } from 'node:path';
+import { fileURLToPath } from 'node:url';
+const MODEL_STEM = 'parakeet-unified-en-0.6b';
+const COREML_FILES = [
+    ['PARAKEET_UNIFIED_COREMLDATA', 'analytics/coremldata.bin'],
+    ['PARAKEET_UNIFIED_COREMLDATA_1', 'coremldata.bin'],
+    ['PARAKEET_UNIFIED_METADATA', 'metadata.json'],
+    ['PARAKEET_UNIFIED_MODEL', 'model.mil'],
+    ['PARAKEET_UNIFIED_WEIGHT', 'weights/weight.bin']
+];
+const defaultAudioPath = fileURLToPath(new URL('../audio/sample-16khz.wav', import.meta.url));
+const modelPath = process.argv[2];
+const audioPath = process.argv[3] ?? defaultAudioPath;
+async function checkModel(path) {
+    const name = basename(path);
+    if (!/^parakeet-unified-en-0\.6b\.(?:f16|q4_0|q8_0)\.gguf$/.test(name)) {
+        throw new Error(`Expected a regenerated Unified GGUF, got: ${name}`);
+    }
+    const file = await open(path, 'r');
+    try {
+        const header = Buffer.alloc(1024 * 1024);
+        const { bytesRead } = await file.read(header, 0, header.length, 0);
+        if (!header.subarray(0, bytesRead).includes('parakeet.unified.left_context_frames')) {
+            throw new Error(`Unified GGUF lacks the new streaming metadata: ${path}`);
+        }
+    }
+    finally {
+        await file.close();
+    }
+}
+async function stageModel(path) {
+    if (process.platform !== 'darwin')
+        throw new Error('Core ML requires macOS');
+    await checkModel(path);
+    const dir = await mkdtemp(join(tmpdir(), 'qvac-unified-coreml-'));
+    try {
+        const modelSrc = join(dir, basename(path));
+        await symlink(resolve(path), modelSrc);
+        const bundleDir = join(dir, `${MODEL_STEM}-encoder.mlmodelc`);
+        for (const [name, relativePath] of COREML_FILES) {
+            const info = await getModelInfo({ name });
+            if (!info.isCached) {
+                if (!info.registrySource || !info.registryPath) {
+                    throw new Error(`Registry location missing for ${name}`);
+                }
+                console.log(`▸ Downloading Core ML ${relativePath}...`);
+                await downloadAsset({ assetSrc: `registry://${info.registrySource}/${info.registryPath}` });
+            }
+            const cached = await getModelInfo({ name });
+            const sourcePath = cached.cacheFiles.find((file) => file.isCached)?.path;
+            if (!sourcePath)
+                throw new Error(`Core ML file is missing from the SDK cache: ${name}`);
+            const targetPath = join(bundleDir, relativePath);
+            await mkdir(join(targetPath, '..'), { recursive: true });
+            await link(sourcePath, targetPath);
+        }
+        return { modelSrc, dir };
+    }
+    catch (error) {
+        await rm(dir, { recursive: true, force: true });
+        throw error;
+    }
+}
+async function main() {
+    if (!modelPath) {
+        throw new Error('Usage: bun run examples/asr/parakeet-unified-coreml.ts ' +
+            '<regenerated-unified-gguf> [16-kHz-wav-file]');
+    }
+    let staged;
+    let modelId;
+    try {
+        staged = await stageModel(modelPath);
+        console.log('▸ Loading Parakeet Unified with its Core ML encoder sidecar...');
+        modelId = await loadModel({
+            modelSrc: staged.modelSrc,
+            modelType: 'parakeet-transcription',
+            modelConfig: { useGPU: true }
+        });
+        console.log('▸ Transcribing audio...');
+        const text = await transcribe({ modelId, audioChunk: audioPath });
+        if (!text.trim())
+            throw new Error('Parakeet Unified returned an empty transcription');
+        console.log(text);
+    }
+    finally {
+        try {
+            if (modelId)
+                await unloadModel({ modelId });
+        }
+        finally {
+            if (staged)
+                await rm(staged.dir, { recursive: true, force: true });
+        }
+    }
+}
+main().catch((error) => {
+    console.error('✖', error);
+    process.exitCode = 1;
+});

package/dist/examples/assess-model-fit.js

@@ -11,8 +11,9 @@
  * call either way. Show it as "can't say", never as "no".
  */
 import { assessModelFit, QWEN3_600M_INST_Q4, QWEN3_1_7B_INST_Q4, QWEN3_4B_INST_Q4_K_M, QWEN3_8B_INST_Q4_K_M, QWEN3_8_27B_MULTIMODAL_UD_Q8_K_XL } from '@qvac/sdk';
-// `as const` so `kind` stays the `'llm'` literal the input type asks for.
-const WORKLOAD = { kind: 'llm', contextTokens: 8192 };
+// Sizes the assessment the way it sizes the load: the estimator reads the
+// context out of the config `loadModel` would be given.
+const CONTEXT_TOKENS = 8192;
 // A ladder ending well past what a laptop has, so one screen shows every verdict.
 const CANDIDATES = [
     QWEN3_600M_INST_Q4,
@@ -31,7 +32,11 @@ function gib(bytes) {
 }
 try {
     const result = await assessModelFit({
-        models: CANDIDATES.map((model) => ({ model, workload: WORKLOAD })),
+        models: CANDIDATES.map((model) => ({
+            modelSrc: model,
+            modelType: 'llamacpp-completion',
+            modelConfig: { ctx_size: CONTEXT_TOKENS }
+        })),
         // Declared for aggregation only, not a scheduling instruction: 'sequential'
         // counts the largest operation peak, 'concurrent' counts one per model.
         execution: 'sequential',
@@ -49,7 +54,7 @@ try {
     const placement = result.assumptions.find((line) => line.includes('assumed to execute on'));
     if (placement)
         console.log(`▸ ${placement}`);
-    console.log(`\n▸ At ${WORKLOAD.contextTokens} tokens of context`);
+    console.log(`\n▸ At ${CONTEXT_TOKENS} tokens of context`);
     for (const [index, model] of result.models.entries()) {
         const mark = VERDICT_MARK[model.verdict];
         const size = gib(CANDIDATES[index].expectedSize).padStart(9);

package/dist/examples/model-fit-ui/server.js

@@ -102,9 +102,12 @@ function sendJson(response, status, payload) {
 async function estimate(body) {
     const names = Array.isArray(body['names']) ? body['names'] : [];
     const contextTokens = Number(body['contextTokens'] ?? 8192);
-    const workload = { kind: 'llm', contextTokens };
     return assessModelFit({
-        models: names.map((name) => ({ model: byName(name), workload })),
+        models: names.map((name) => ({
+            modelSrc: byName(name),
+            modelType: 'llamacpp-completion',
+            modelConfig: { ctx_size: contextTokens }
+        })),
         // Declared for aggregation only. 'sequential' counts the largest operation
         // peak; 'concurrent' counts one per model.
         execution: 'sequential',

package/dist/examples/rpc-inference.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=rpc-inference.d.ts.map