@missionsquad/browserai 0.1.0

package/dist/models.js

@@ -0,0 +1,1052 @@
+/** Friendly capability names surfaced across the app (type chips/filters, setup dropdown/filters). */
+export const TASK_LABELS = { text: "Text", vision: "Image", stt: "Transcription", tts: "Speech" };
+export const TASK_ICONS = {
+    text: "mdi-text-box-outline",
+    vision: "mdi-image-outline",
+    stt: "mdi-microphone-outline",
+    tts: "mdi-account-voice",
+};
+export const MODALITY_LABELS = { text: "Text", image: "Image", audio: "Audio" };
+export const MODALITY_ICONS = { text: "mdi-text", image: "mdi-image-outline", audio: "mdi-waveform" };
+/** Task capabilities for a preset, defaulting a bare LLM to ["text"]. */
+export function presetTasks(preset) {
+    return preset.tasks && preset.tasks.length > 0 ? preset.tasks : ["text"];
+}
+/** Canonical ordering of the four capability slots (Text / Image / Transcription / Speech). */
+export const SLOT_ORDER = ["text", "vision", "stt", "tts"];
+/**
+ * The capability SLOTS a model occupies when loaded — one per its {@link presetTasks}. Used to bound
+ * concurrent loads to one model per slot (Text/Image/Transcription/Speech).
+ *
+ * NOTE: image-text-to-text VLMs (SmolVLM/Qwen3-VL/GLM-OCR) do NOT fill the Text slot — their
+ * Transformers.js processors require an image and error ("No images provided") on text-only input, so
+ * they cannot drive the text tabs. Only genuine text runtimes (LLMs, and Gemma-4 multimodal which has
+ * a verified text path) fill Text. A VLM therefore occupies just the Image slot.
+ */
+export function modelSlots(preset) {
+    const slots = new Set(presetTasks(preset));
+    return SLOT_ORDER.filter((task) => slots.has(task));
+}
+export function presetModalityIn(preset) {
+    return preset.modalityIn && preset.modalityIn.length > 0 ? preset.modalityIn : ["text"];
+}
+export function presetModalityOut(preset) {
+    return preset.modalityOut && preset.modalityOut.length > 0 ? preset.modalityOut : ["text"];
+}
+/** Model ids of the TTS presets, used to look up their voice rosters. */
+export const KOKORO_MODEL_ID = "onnx-community/Kokoro-82M-v1.0-ONNX";
+export const SUPERTONIC_MODEL_ID = "onnx-community/Supertonic-TTS-ONNX";
+export const SUPERTONIC_2_MODEL_ID = "onnx-community/Supertonic-TTS-2-ONNX";
+/** Kokoro voice roster surfaced in the TTS tab (id → display name + tag), verified voice ids. */
+export const KOKORO_VOICES = [
+    { id: "af_heart", label: "Heart", tag: "US female" },
+    { id: "af_bella", label: "Bella", tag: "US female" },
+    { id: "am_michael", label: "Michael", tag: "US male" },
+    { id: "am_fenrir", label: "Fenrir", tag: "US male" },
+    { id: "bf_emma", label: "Emma", tag: "UK female" },
+    { id: "bm_george", label: "George", tag: "UK male" },
+];
+/**
+ * Supertonic speaker roster. The repo ships ten speaker-embedding files under `voices/` — F1–F5
+ * (female) and M1–M5 (male) — verified against the Hugging Face repo file tree. The voice id is the
+ * `.bin` basename consumed as the pipeline's `speaker_embeddings`.
+ */
+export const SUPERTONIC_VOICES = [
+    { id: "F1", label: "Female 1", tag: "English" },
+    { id: "F2", label: "Female 2", tag: "English" },
+    { id: "F3", label: "Female 3", tag: "English" },
+    { id: "F4", label: "Female 4", tag: "English" },
+    { id: "F5", label: "Female 5", tag: "English" },
+    { id: "M1", label: "Male 1", tag: "English" },
+    { id: "M2", label: "Male 2", tag: "English" },
+    { id: "M3", label: "Male 3", tag: "English" },
+    { id: "M4", label: "Male 4", tag: "English" },
+    { id: "M5", label: "Male 5", tag: "English" },
+];
+/** The selectable voices for a TTS model (empty for non-TTS models). */
+export function ttsVoicesForModel(modelId) {
+    if (modelId === KOKORO_MODEL_ID)
+        return KOKORO_VOICES;
+    if (modelId === SUPERTONIC_MODEL_ID || modelId === SUPERTONIC_2_MODEL_ID)
+        return SUPERTONIC_VOICES;
+    return [];
+}
+function extractionPreset(overrides = {}) {
+    return {
+        maxTokens: 512,
+        temperature: 0,
+        topP: 0.85,
+        repetitionPenalty: 1.08,
+        frequencyPenalty: 0.2,
+        presencePenalty: 0,
+        seed: 42,
+        disableThinking: false,
+        ...overrides,
+    };
+}
+const gemmaFixedContextOverrides = {
+    // Gemma 3 and current experimental Gemma 4 MLC configs can expose both values as positive
+    // (for example context_window_size=4096 and sliding_window_size=512). WebLLM requires one
+    // KV-cache strategy at load time, so use the fixed context-window path for this demo.
+    context_window_size: 4096,
+    sliding_window_size: -1,
+    max_history_size: 1,
+};
+export const MODEL_RECORD_OVERRIDES = {
+    "gemma3-1b-it-q4f16_1-MLC": gemmaFixedContextOverrides,
+    "gemma-4-E2B-it-q4f16_1-MLC": gemmaFixedContextOverrides,
+    "gemma-4-E4B-it-q4f16_1-MLC": gemmaFixedContextOverrides,
+};
+export function applyCuratedModelRecordOverrides(record) {
+    const overrides = MODEL_RECORD_OVERRIDES[record.model_id];
+    if (!overrides)
+        return record;
+    return {
+        ...record,
+        overrides: {
+            ...(record.overrides ?? {}),
+            ...overrides,
+        },
+    };
+}
+// Gemma 4 is not bundled in @mlc-ai/web-llm 0.2.84's prebuiltAppConfig.
+// WebLLM supports custom MLC model records, so these third-party records are appended
+// to the runtime AppConfig and their artifacts are proxied through the Cloudflare Worker.
+// Keep these marked experimental until they are tested on your target hardware.
+export const CUSTOM_MODEL_RECORDS = [
+    {
+        model: "https://huggingface.co/welcoma/gemma-4-E2B-it-q4f16_1-MLC",
+        model_id: "gemma-4-E2B-it-q4f16_1-MLC",
+        model_lib: "https://huggingface.co/welcoma/gemma-4-E2B-it-q4f16_1-MLC/resolve/main/libs/gemma-4-E2B-it-q4f16_1-MLC-webgpu.wasm",
+        vram_required_MB: 4096,
+        low_resource_required: false,
+        required_features: ["shader-f16"],
+        overrides: gemmaFixedContextOverrides,
+    },
+    {
+        model: "https://huggingface.co/welcoma/gemma-4-E4B-it-q4f16_1-MLC",
+        model_id: "gemma-4-E4B-it-q4f16_1-MLC",
+        model_lib: "https://huggingface.co/welcoma/gemma-4-E4B-it-q4f16_1-MLC/resolve/main/libs/gemma-4-E4B-it-q4f16_1-MLC-webgpu.wasm",
+        vram_required_MB: 6144,
+        low_resource_required: false,
+        required_features: ["shader-f16"],
+        overrides: gemmaFixedContextOverrides,
+    },
+];
+export const CUSTOM_MODEL_RECORD_IDS = new Set(CUSTOM_MODEL_RECORDS.map((record) => record.model_id));
+export const MODEL_PRESETS = [
+    {
+        id: "Qwen3.5-0.8B-q4f16_1-MLC",
+        label: "Qwen3.5 0.8B q4f16 — recommended tested default (~1.6 GB VRAM)",
+        shortName: "Qwen3.5 0.8B",
+        description: "Default browser-extraction model. It is the smallest recent WebLLM Qwen3.5 option and has worked well for deterministic JSON extraction in this app.",
+        approximateDownload: "~447 MB model files plus browser cache overhead",
+        vramRequiredMB: 1629.49,
+        vramSource: "webllm",
+        lowResourceRequired: true,
+        disableThinkingSupported: true,
+        defaultRuntime: extractionPreset({ disableThinking: true }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "recommended",
+        officialRepo: "https://huggingface.co/mlc-ai/Qwen3.5-0.8B-q4f16_1-MLC",
+    },
+    {
+        id: "Qwen3.5-2B-q4f16_1-MLC",
+        label: "Qwen3.5 2B q4f16 — better quality, moderate memory (~2.2 GB VRAM)",
+        shortName: "Qwen3.5 2B",
+        description: "Larger Qwen3.5 option for better instruction following and JSON extraction when the device has enough GPU memory.",
+        approximateDownload: "larger multi-file model download plus browser cache overhead",
+        vramRequiredMB: 2245.44,
+        vramSource: "webllm",
+        lowResourceRequired: false,
+        disableThinkingSupported: true,
+        defaultRuntime: extractionPreset({ maxTokens: 640, disableThinking: true }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/Qwen3.5-2B-q4f16_1-MLC",
+    },
+    {
+        id: "Qwen3.5-4B-q4f16_1-MLC",
+        label: "Qwen3.5 4B q4f16 — higher quality desktop option (~3.8 GB VRAM)",
+        shortName: "Qwen3.5 4B",
+        description: "Higher-quality Qwen3.5 option for desktop GPUs. Use when extraction quality is more important than load time and memory footprint.",
+        approximateDownload: "large multi-file model download plus browser cache overhead",
+        vramRequiredMB: 3867.82,
+        vramSource: "webllm",
+        lowResourceRequired: false,
+        disableThinkingSupported: true,
+        defaultRuntime: extractionPreset({ maxTokens: 768, disableThinking: true, repetitionPenalty: 1.06 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/Qwen3.5-4B-q4f16_1-MLC",
+    },
+    {
+        id: "gemma3-1b-it-q4f16_1-MLC",
+        label: "Gemma 3 1B IT q4f16 — small Google fallback (~711 MB VRAM)",
+        shortName: "Gemma 3 1B",
+        description: "Small official WebLLM Gemma 3 model. This build uses a fixed-context override in the app so WebLLM does not try to enable both context-window and sliding-window KV-cache modes.",
+        approximateDownload: "small model download plus browser cache overhead",
+        vramRequiredMB: 711.07,
+        vramSource: "webllm",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 512, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/gemma3-1b-it-q4f16_1-MLC",
+    },
+    {
+        id: "gemma-4-E2B-it-q4f16_1-MLC",
+        label: "Gemma 4 E2B IT q4f16 — experimental custom WebLLM build (~4 GB VRAM est.)",
+        shortName: "Gemma 4 E2B",
+        description: "Smallest Gemma 4 instruction option available here. This uses a third-party custom MLC/WebLLM artifact and a fixed-context override for WebLLM KV-cache loading.",
+        approximateDownload: "~2.7 GB model repo plus browser cache overhead",
+        vramRequiredMB: 4096,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "custom-mlc",
+        backend: "webllm",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/google/gemma-4-E2B-it",
+        artifactRepo: "https://huggingface.co/welcoma/gemma-4-E2B-it-q4f16_1-MLC",
+    },
+    {
+        id: "gemma-4-E4B-it-q4f16_1-MLC",
+        label: "Gemma 4 E4B IT q4f16 — experimental larger Gemma 4 build (~6 GB VRAM est.)",
+        shortName: "Gemma 4 E4B",
+        description: "Larger Gemma 4 instruction option for desktop/high-memory testing. This uses the same fixed-context WebLLM load override as the E2B option.",
+        approximateDownload: "~4.3 GB model repo plus browser cache overhead",
+        vramRequiredMB: 6144,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "custom-mlc",
+        backend: "webllm",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/google/gemma-4-E4B-it",
+        artifactRepo: "https://huggingface.co/welcoma/gemma-4-E4B-it-q4f16_1-MLC",
+    },
+    {
+        id: "onnx-community/whisper-base",
+        label: "Whisper Base — multilingual speech-to-text via Transformers.js/WebGPU",
+        shortName: "Whisper Base",
+        description: "OpenAI Whisper base — dependable multilingual speech-to-text with punctuation and segment timestamps.",
+        approximateDownload: "~145 MB (encoder fp32 + q4/q8 decoder) plus browser cache overhead",
+        vramRequiredMB: 500,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "recommended",
+        officialRepo: "https://huggingface.co/onnx-community/whisper-base",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "stt",
+        brand: "OpenAI",
+    },
+    {
+        id: "onnx-community/whisper-tiny",
+        label: "Whisper Tiny — smallest multilingual speech-to-text",
+        shortName: "Whisper Tiny",
+        description: "Smallest Whisper. Ultra-fast multilingual transcription for low-memory devices.",
+        approximateDownload: "~41 MB q8 ONNX plus browser cache overhead",
+        vramRequiredMB: 250,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/onnx-community/whisper-tiny",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "stt",
+        brand: "OpenAI",
+    },
+    {
+        id: "onnx-community/moonshine-base-ONNX",
+        label: "Moonshine Base — fast English on-device speech-to-text",
+        shortName: "Moonshine Base",
+        description: "Fast English speech-to-text tuned for short, real-time clips and low-latency on-device transcription.",
+        approximateDownload: "~63 MB q8 ONNX plus browser cache overhead",
+        vramRequiredMB: 400,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/onnx-community/moonshine-base-ONNX",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "stt",
+        brand: "Moonshine",
+    },
+    {
+        id: "onnx-community/moonshine-tiny-ONNX",
+        label: "Moonshine Tiny — smallest fast English on-device speech-to-text",
+        shortName: "Moonshine Tiny",
+        description: "Smallest Moonshine. Ultra-low-latency English speech-to-text for short, real-time clips on low-memory devices.",
+        approximateDownload: "~75 MB (fp32 encoder + q4 decoder) plus browser cache overhead",
+        vramRequiredMB: 250,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/onnx-community/moonshine-tiny-ONNX",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "stt",
+        brand: "Moonshine",
+    },
+    {
+        id: "onnx-community/whisper-medium_timestamped",
+        label: "Whisper Medium — higher-accuracy multilingual speech-to-text (word timestamps)",
+        shortName: "Whisper Medium",
+        description: "OpenAI Whisper medium (~769M) with word-level timestamps. Stronger multilingual accuracy than Base/Tiny, at a much larger download and slower first load.",
+        approximateDownload: "~1.7 GB (fp32 encoder + q4 decoder) plus browser cache overhead",
+        vramRequiredMB: 1800,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/onnx-community/whisper-medium_timestamped",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "stt",
+        brand: "OpenAI",
+    },
+    {
+        id: "onnx-community/whisper-medium.en_timestamped",
+        label: "Whisper Medium English — English-only medium speech-to-text (word timestamps)",
+        shortName: "Whisper Medium EN",
+        description: "English-only Whisper medium (~769M) with word-level timestamps. Tuned for English accuracy; do not use it for other languages.",
+        approximateDownload: "~1.7 GB (fp32 encoder + q4 decoder) plus browser cache overhead",
+        vramRequiredMB: 1800,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/onnx-community/whisper-medium.en_timestamped",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "stt",
+        brand: "OpenAI",
+    },
+    {
+        id: "onnx-community/cohere-transcribe-03-2026-ONNX",
+        label: "Cohere Transcribe — multilingual speech-to-text via Transformers.js/WebGPU",
+        shortName: "Cohere Transcribe",
+        description: "Cohere's multilingual transcription model (14 languages). Loads via the automatic-speech-recognition pipeline with a flat q4 dtype — its fp32 encoder is multiple GB, so q4 is the documented WebGPU recipe.",
+        approximateDownload: "~2.1 GB q4 ONNX (q4 encoder + q4 decoder) plus browser cache overhead",
+        vramRequiredMB: 2600,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/onnx-community/cohere-transcribe-03-2026-ONNX",
+        transformersDtype: "q4",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "stt",
+        brand: "Cohere",
+    },
+    {
+        id: "onnx-community/parakeet-ctc-0.6b-ONNX",
+        label: "Parakeet CTC 0.6B — fast low-latency English speech-to-text (NVIDIA FastConformer)",
+        shortName: "Parakeet CTC 0.6B",
+        description: "NVIDIA Parakeet CTC 0.6B (FastConformer). Fast, low-latency English transcription through the CTC speech-recognition pipeline. Loads with a flat q4f16 dtype (single-module CTC model — no separate encoder/decoder).",
+        approximateDownload: "~450 MB q4f16 ONNX plus browser cache overhead",
+        vramRequiredMB: 1200,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/onnx-community/parakeet-ctc-0.6b-ONNX",
+        transformersDtype: "q4f16",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "stt",
+    },
+    {
+        id: "onnx-community/granite-speech-4.1-2b-ONNX",
+        label: "IBM Granite Speech 4.1 2B — audio-LLM speech-to-text via Transformers.js/WebGPU",
+        shortName: "Granite Speech 2B",
+        description: "IBM Granite Speech (2B): a speech-aware LLM (Conformer encoder + Granite LLM + audio LoRA) with strong English ASR. Runs through the audio-text-to-text recipe — it generates the transcript, so there are no per-segment timestamps.",
+        approximateDownload: "~1.5 GB q4f16 ONNX (audio encoder + decoder + embeddings) plus browser cache overhead",
+        vramRequiredMB: 2200,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/onnx-community/granite-speech-4.1-2b-ONNX",
+        transformersDtype: "q4f16",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "audio-text-to-text",
+        brand: "IBM Granite",
+    },
+    {
+        id: "onnx-community/Voxtral-Mini-4B-Realtime-2602-ONNX",
+        label: "Voxtral Mini 4B Realtime — Mistral audio-LLM speech-to-text via Transformers.js/WebGPU",
+        shortName: "Voxtral Mini 4B RT",
+        description: "Mistral Voxtral Mini 4B: an audio-LLM for multilingual transcription (14 languages) built on Ministral. Runs through the audio-text-to-text recipe as a single-pass transcription. This is the realtime/streaming checkpoint and is heavy (4B), so single-pass output quality should be validated on your hardware.",
+        approximateDownload: "~2.8 GB q4f16 ONNX (audio encoder + decoder + embeddings) plus browser cache overhead",
+        vramRequiredMB: 4096,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602",
+        artifactRepo: "https://huggingface.co/onnx-community/Voxtral-Mini-4B-Realtime-2602-ONNX",
+        transformersDtype: "q4f16",
+        tasks: ["stt"],
+        modalityIn: ["audio"],
+        modalityOut: ["text"],
+        mmRuntime: "voxtral-realtime",
+    },
+    {
+        id: "onnx-community/Kokoro-82M-v1.0-ONNX",
+        label: "Kokoro 82M — natural on-device text-to-speech (kokoro-js)",
+        shortName: "Kokoro 82M",
+        description: "Compact, natural text-to-speech with a library of expressive voices. A browser-TTS favorite.",
+        approximateDownload: "~86 MB q8 (326 MB fp32 on WebGPU) plus browser cache overhead",
+        vramRequiredMB: 420,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "recommended",
+        officialRepo: "https://huggingface.co/onnx-community/Kokoro-82M-v1.0-ONNX",
+        tasks: ["tts"],
+        modalityIn: ["text"],
+        modalityOut: ["audio"],
+        mmRuntime: "tts-kokoro",
+        brand: "Kokoro",
+        notes: "Loaded via kokoro-js from CDN; it bundles its own Transformers.js runtime and fetches weights directly from Hugging Face.",
+    },
+    {
+        id: "onnx-community/Supertonic-TTS-ONNX",
+        label: "Supertonic — low-latency streaming text-to-speech (native pipeline)",
+        shortName: "Supertonic",
+        description: "Very low-latency on-device text-to-speech built for real-time streaming playback.",
+        approximateDownload: "~fp32 ONNX plus browser cache overhead",
+        vramRequiredMB: 350,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/onnx-community/Supertonic-TTS-ONNX",
+        tasks: ["tts"],
+        modalityIn: ["text"],
+        modalityOut: ["audio"],
+        mmRuntime: "tts-pipeline",
+        brand: "Supertonic",
+    },
+    {
+        id: "onnx-community/Supertonic-TTS-2-ONNX",
+        label: "Supertonic 2 — multilingual low-latency streaming text-to-speech (native pipeline)",
+        shortName: "Supertonic 2",
+        description: "Second-generation Supertonic: very low-latency on-device text-to-speech for real-time streaming playback, now multilingual (English, Korean, Spanish, Portuguese, French). Same ten F1–F5 / M1–M5 speaker voices as v1.",
+        approximateDownload: "~260 MB fp32 ONNX plus browser cache overhead",
+        vramRequiredMB: 400,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/onnx-community/Supertonic-TTS-2-ONNX",
+        tasks: ["tts"],
+        modalityIn: ["text"],
+        modalityOut: ["audio"],
+        mmRuntime: "tts-pipeline",
+        brand: "Supertonic",
+    },
+    {
+        id: "HuggingFaceTB/SmolVLM-256M-Instruct",
+        label: "SmolVLM 256M — tiny vision-language model via Transformers.js/WebGPU",
+        shortName: "SmolVLM 256M",
+        description: "Tiny vision-language model for captioning and quick image Q&A. Runs on almost any device.",
+        approximateDownload: "~190 MB q4f16 ONNX plus browser cache overhead",
+        vramRequiredMB: 600,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/HuggingFaceTB/SmolVLM-256M-Instruct",
+        transformersDtype: "q4",
+        tasks: ["vision"],
+        modalityIn: ["text", "image"],
+        modalityOut: ["text"],
+        mmRuntime: "vlm-vision2seq",
+    },
+    {
+        id: "HuggingFaceTB/SmolVLM-500M-Instruct",
+        label: "SmolVLM 500M — stronger small vision-language model",
+        shortName: "SmolVLM 500M",
+        description: "Small vision-language model with stronger scene and document understanding than the 256M build.",
+        approximateDownload: "~250 MB q4f16 ONNX plus browser cache overhead",
+        vramRequiredMB: 900,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/HuggingFaceTB/SmolVLM-500M-Instruct",
+        transformersDtype: "q4",
+        tasks: ["vision"],
+        modalityIn: ["text", "image"],
+        modalityOut: ["text"],
+        mmRuntime: "vlm-vision2seq",
+    },
+    {
+        id: "onnx-community/Qwen3-VL-2B-Instruct-ONNX",
+        label: "Qwen3-VL 2B — capable vision-language model via Transformers.js/WebGPU",
+        shortName: "Qwen3-VL 2B",
+        description: "Capable vision-language model for detailed image reasoning, charts, screenshots, and multi-step questions.",
+        approximateDownload: "~1.5 GB mixed-precision ONNX plus browser cache overhead",
+        vramRequiredMB: 2500,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/Qwen/Qwen3-VL-2B-Instruct",
+        artifactRepo: "https://huggingface.co/onnx-community/Qwen3-VL-2B-Instruct-ONNX",
+        tasks: ["vision"],
+        modalityIn: ["text", "image"],
+        modalityOut: ["text"],
+        mmRuntime: "vlm-imagetext",
+    },
+    {
+        id: "wolfofbackstreet/GLM-OCR-ONNX-q4f16",
+        label: "GLM-OCR — document OCR vision-language model via Transformers.js/WebGPU",
+        shortName: "GLM-OCR",
+        description: "Document-focused vision model for OCR, layout parsing, and structured text extraction from images.",
+        approximateDownload: "~658 MB q4f16 ONNX plus browser cache overhead",
+        vramRequiredMB: 1800,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/zai-org/GLM-OCR",
+        artifactRepo: "https://huggingface.co/wolfofbackstreet/GLM-OCR-ONNX-q4f16",
+        transformersDtype: "q4f16",
+        tasks: ["vision"],
+        modalityIn: ["text", "image"],
+        modalityOut: ["text"],
+        mmRuntime: "vlm-imagetext",
+        brand: "Zhipu",
+    },
+    {
+        id: "onnx-community/Qwen2.5-VL-3B-Instruct-ONNX",
+        label: "Qwen2.5-VL 3B — vision-language model via Transformers.js/WebGPU",
+        shortName: "Qwen2.5-VL 3B",
+        description: "Qwen2.5-VL 3B for detailed image reasoning, charts, screenshots, and documents. Larger and stronger than the Qwen3-VL 2B build, at a heavier first download and higher GPU-memory needs.",
+        approximateDownload: "~3.7 GB mixed-precision ONNX (fp16 vision + q4f16 decoder) plus browser cache overhead",
+        vramRequiredMB: 3500,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/Qwen/Qwen2.5-VL-3B-Instruct",
+        artifactRepo: "https://huggingface.co/onnx-community/Qwen2.5-VL-3B-Instruct-ONNX",
+        tasks: ["vision"],
+        modalityIn: ["text", "image"],
+        modalityOut: ["text"],
+        mmRuntime: "vlm-imagetext",
+    },
+    {
+        id: "onnx-community/LightOnOCR-2-1B-ONNX",
+        label: "LightOnOCR-2 1B — document OCR vision-language model via Transformers.js/WebGPU",
+        shortName: "LightOnOCR-2 1B",
+        description: "Compact document-OCR vision model for text extraction, tables, forms, and layout parsing across 11 languages. Loads flat q4f16 like the GLM-OCR build.",
+        approximateDownload: "~680 MB q4f16 ONNX plus browser cache overhead",
+        vramRequiredMB: 1500,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset(),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/lightonai/LightOnOCR-2-1B",
+        artifactRepo: "https://huggingface.co/onnx-community/LightOnOCR-2-1B-ONNX",
+        transformersDtype: "q4f16",
+        tasks: ["vision"],
+        modalityIn: ["text", "image"],
+        modalityOut: ["text"],
+        mmRuntime: "vlm-imagetext",
+        brand: "LightOn",
+    },
+    {
+        id: "onnx-community/NVIDIA-Nemotron-3-Nano-4B-BF16-ONNX",
+        label: "NVIDIA Nemotron 3 Nano 4B ONNX — experimental Transformers.js/WebGPU (~2.5 GB download)",
+        shortName: "Nemotron 3 Nano 4B",
+        description: "Experimental ONNX + Transformers.js backend from the WebGPU demo space. It is prompt-only JSON generation, not WebLLM schema-constrained decoding, and may require substantially more disk/GPU memory than Qwen3.5.",
+        approximateDownload: "~2.5 GB browser-selected quantized ONNX artifacts, cached after first load",
+        vramRequiredMB: 4096,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: true,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.95, repetitionPenalty: 1.05, disableThinking: true }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16",
+        artifactRepo: "https://huggingface.co/onnx-community/NVIDIA-Nemotron-3-Nano-4B-BF16-ONNX",
+        transformersDtype: "q4",
+        notes: "Loaded through @huggingface/transformers from CDN at runtime so the WebLLM demo remains small until this model is selected.",
+    },
+    {
+        id: "LiquidAI/LFM2.5-230M-ONNX",
+        label: "Liquid LFM2.5 230M ONNX — tiny Transformers.js/WebGPU model (~300 MB class)",
+        shortName: "LFM2.5 230M",
+        description: "Very small LiquidAI ONNX model for browser testing. It is useful when you want a fast, low-footprint Transformers.js/WebGPU baseline for extraction prompts.",
+        approximateDownload: "~300 MB class model artifacts plus browser cache overhead",
+        vramRequiredMB: 768,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 384, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/LiquidAI/LFM2.5-230M",
+        artifactRepo: "https://huggingface.co/LiquidAI/LFM2.5-230M-ONNX",
+        transformersDtype: "q4",
+    },
+    {
+        id: "LiquidAI/LFM2.5-350M-ONNX",
+        label: "Liquid LFM2.5 350M ONNX — compact Transformers.js/WebGPU model (~700 MB class)",
+        shortName: "LFM2.5 350M",
+        description: "Compact LiquidAI edge model with an official ONNX/WebGPU export. Good for comparing a hybrid architecture against Qwen, Gemma, and Llama in browser extraction.",
+        approximateDownload: "~700 MB class model artifacts plus browser cache overhead",
+        vramRequiredMB: 1024,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 448, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/LiquidAI/LFM2.5-350M",
+        artifactRepo: "https://huggingface.co/LiquidAI/LFM2.5-350M-ONNX",
+        transformersDtype: "q4",
+    },
+    {
+        id: "LiquidAI/LFM2.5-1.2B-Thinking-ONNX",
+        label: "Liquid LFM2.5 1.2B Thinking ONNX — reasoning-focused browser model",
+        shortName: "LFM2.5 1.2B Thinking",
+        description: "Reasoning-focused LiquidAI ONNX model for Transformers.js/WebGPU. It may emit <think> traces, so extraction prompts should explicitly request final JSON only.",
+        approximateDownload: "large ONNX model artifacts plus browser cache overhead",
+        vramRequiredMB: 2300,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: true,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.06, disableThinking: true }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking",
+        artifactRepo: "https://huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking-ONNX",
+        transformersDtype: "q4",
+    },
+    {
+        id: "onnx-community/Falcon-H1-Tiny-90M-Instruct-ONNX",
+        label: "Falcon H1 Tiny 90M ONNX — ultra-small Transformers.js/WebGPU smoke-test model",
+        shortName: "Falcon H1 Tiny 90M",
+        description: "Ultra-small hybrid Mamba/attention model. It is best as a browser compatibility smoke test; extraction quality will likely be lower than the larger models.",
+        approximateDownload: "~1.1 GB model repo, but selected quantized artifacts may be smaller",
+        vramRequiredMB: 512,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 256, topP: 0.9, repetitionPenalty: 1.04 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/tiiuae/Falcon-H1-Tiny-90M-Instruct",
+        artifactRepo: "https://huggingface.co/onnx-community/Falcon-H1-Tiny-90M-Instruct-ONNX",
+        transformersDtype: "q4",
+    },
+    {
+        id: "onnx-community/granite-4.0-micro-ONNX-web",
+        label: "IBM Granite 4.0 Micro ONNX-web — small hybrid chat model via Transformers.js/WebGPU",
+        shortName: "Granite 4.0 Micro ONNX",
+        description: "Small IBM Granite MoE/hybrid chat model packaged specifically for Transformers.js/WebGPU. Useful as a non-Qwen/non-Llama browser extraction comparison.",
+        approximateDownload: "ONNX-web model artifacts plus tokenizer/config files",
+        vramRequiredMB: 2048,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 512, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/ibm-granite/granite-4.0-micro",
+        artifactRepo: "https://huggingface.co/onnx-community/granite-4.0-micro-ONNX-web",
+        transformersDtype: "q4f16",
+    },
+    {
+        id: "onnx-community/granite-4.0-350m-ONNX-web",
+        label: "IBM Granite 4.0 350M ONNX-web — small Transformers.js/WebGPU model",
+        shortName: "Granite 4.0 350M",
+        description: "Small Granite 4.0 ONNX-web build packaged for Transformers.js/WebGPU. This is one of the safest Granite test options for lower-memory machines.",
+        approximateDownload: "~350 MB q4f16 ONNX artifacts plus tokenizer/config files",
+        vramRequiredMB: 1024,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 448, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/ibm-granite/granite-4.0-350m",
+        artifactRepo: "https://huggingface.co/onnx-community/granite-4.0-350m-ONNX-web",
+        transformersDtype: "q4f16",
+    },
+    {
+        id: "onnx-community/granite-4.0-h-350m-ONNX",
+        label: "IBM Granite 4.0 H 350M ONNX — hybrid/MoE Transformers.js/WebGPU test",
+        shortName: "Granite 4.0 H 350M",
+        description: "Small Granite hybrid/MoE ONNX export. Use it to compare the Granite H family against Qwen3.5 and LiquidAI in browser extraction.",
+        approximateDownload: "~236 MB q4f16 ONNX data plus graph/tokenizer/config files",
+        vramRequiredMB: 1024,
+        vramSource: "estimated",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 448, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/ibm-granite/granite-4.0-h-350m",
+        artifactRepo: "https://huggingface.co/onnx-community/granite-4.0-h-350m-ONNX",
+        transformersDtype: "q4f16",
+    },
+    {
+        id: "onnx-community/granite-4.0-1b-ONNX-web",
+        label: "IBM Granite 4.0 1B ONNX-web — larger Transformers.js/WebGPU model",
+        shortName: "Granite 4.0 1B",
+        description: "Granite 4.0 1B ONNX-web build packaged for Transformers.js/WebGPU. Expect better extraction quality than the 350M variant and a larger first download.",
+        approximateDownload: "~1.25 GB q4f16 ONNX artifacts plus tokenizer/config files",
+        vramRequiredMB: 2048,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 640, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/ibm-granite/granite-4.0-1b",
+        artifactRepo: "https://huggingface.co/onnx-community/granite-4.0-1b-ONNX-web",
+        transformersDtype: "q4f16",
+    },
+    {
+        id: "onnx-community/granite-4.0-h-1b-ONNX",
+        label: "IBM Granite 4.0 H 1B ONNX — hybrid/MoE Transformers.js/WebGPU model",
+        shortName: "Granite 4.0 H 1B",
+        description: "Larger Granite H hybrid/MoE ONNX export. Use this on machines that can handle around a 1 GB quantized ONNX download plus runtime memory.",
+        approximateDownload: "~925 MB q4f16 ONNX data plus graph/tokenizer/config files",
+        vramRequiredMB: 2048,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 640, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/ibm-granite/granite-4.0-h-1b",
+        artifactRepo: "https://huggingface.co/onnx-community/granite-4.0-h-1b-ONNX",
+        transformersDtype: "q4f16",
+    },
+    {
+        id: "onnx-community/granite-4.0-h-micro-ONNX",
+        label: "IBM Granite 4.0 H Micro ONNX — larger hybrid/MoE test",
+        shortName: "Granite 4.0 H Micro",
+        description: "Large Granite H Micro ONNX export. It is included for testing but is not a first-load choice because the repository contains very large model variants.",
+        approximateDownload: "~1.95 GB q4f16 ONNX data plus graph/tokenizer/config files",
+        vramRequiredMB: 4096,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/ibm-granite/granite-4.0-h-micro",
+        artifactRepo: "https://huggingface.co/onnx-community/granite-4.0-h-micro-ONNX",
+        transformersDtype: "q4f16",
+    },
+    {
+        id: "Mike0021/MiniCPM5-1B-ONNX-Web",
+        label: "MiniCPM5 1B ONNX-web — community q4 Transformers.js/WebGPU export",
+        shortName: "MiniCPM5 1B ONNX-web",
+        description: "Browser-friendly q4 Transformers.js export of MiniCPM5 1B. The official ONNX Runtime GenAI repo is linked as the official repo; this selected artifact repo is the browser-loadable q4 export.",
+        approximateDownload: "~902 MB q4 ONNX artifact plus tokenizer/config files",
+        vramRequiredMB: 2048,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: true,
+        defaultRuntime: extractionPreset({ maxTokens: 640, topP: 0.9, repetitionPenalty: 1.05, disableThinking: true }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/onnx-community/MiniCPM5-1B",
+        artifactRepo: "https://huggingface.co/Mike0021/MiniCPM5-1B-ONNX-Web",
+        transformersDtype: "q4",
+    },
+    {
+        id: "HuggingFaceTB/SmolLM3-3B-ONNX",
+        label: "SmolLM3 3B ONNX — current compact 3B model via Transformers.js/WebGPU",
+        shortName: "SmolLM3 3B ONNX",
+        description: "Current SmolLM3 ONNX export for Transformers.js. It is larger than the default model but useful for testing a fully open 3B browser/local model path.",
+        approximateDownload: "large ONNX model artifacts plus tokenizer/config files",
+        vramRequiredMB: 3072,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/HuggingFaceTB/SmolLM3-3B",
+        artifactRepo: "https://huggingface.co/HuggingFaceTB/SmolLM3-3B-ONNX",
+        transformersDtype: "q4",
+    },
+    {
+        id: "onnx-community/gemma-4-E2B-it-ONNX",
+        label: "Gemma 4 E2B ONNX — experimental Transformers.js/WebGPU alternative",
+        shortName: "Gemma 4 E2B ONNX",
+        description: "ONNX/Transformers.js route for Gemma 4 E2B. This is separate from the custom WebLLM/MLC Gemma 4 build and may be useful if the MLC artifact is not stable on a device.",
+        approximateDownload: "large ONNX model artifacts plus tokenizer/config files",
+        vramRequiredMB: 4096,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/google/gemma-4-E2B-it",
+        artifactRepo: "https://huggingface.co/onnx-community/gemma-4-E2B-it-ONNX",
+        transformersDtype: "q4f16",
+        tasks: ["text", "vision", "stt"],
+        modalityIn: ["text", "image", "audio"],
+        modalityOut: ["text"],
+        mmRuntime: "gemma-mm",
+    },
+    {
+        id: "onnx-community/gemma-4-E4B-it-ONNX",
+        label: "Gemma 4 E4B ONNX — larger multimodal Transformers.js/WebGPU build",
+        shortName: "Gemma 4 E4B ONNX",
+        description: "Larger ONNX/Transformers.js route for Gemma 4 E4B (~8B). Text + image + audio in one model, with stronger quality than the E2B ONNX build at a heavier first load and higher GPU-memory needs.",
+        approximateDownload: "~5–6 GB q4f16 ONNX artifacts plus tokenizer/config files",
+        vramRequiredMB: 6144,
+        vramSource: "estimated",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "transformers-onnx",
+        backend: "transformers-js",
+        stability: "experimental",
+        officialRepo: "https://huggingface.co/google/gemma-4-E4B-it",
+        artifactRepo: "https://huggingface.co/onnx-community/gemma-4-E4B-it-ONNX",
+        transformersDtype: "q4f16",
+        tasks: ["text", "vision", "stt"],
+        modalityIn: ["text", "image", "audio"],
+        modalityOut: ["text"],
+        mmRuntime: "gemma-mm",
+    },
+    {
+        id: "Llama-3.2-1B-Instruct-q4f16_1-MLC",
+        label: "Llama 3.2 1B Instruct q4f16 — low-memory Meta fallback (~879 MB VRAM)",
+        shortName: "Llama 3.2 1B",
+        description: "Compact Meta Llama instruction model. It is a useful compatibility fallback and gives you another model family to compare against Qwen3.5.",
+        approximateDownload: "small model download plus browser cache overhead",
+        vramRequiredMB: 879.04,
+        vramSource: "webllm",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 448, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/Llama-3.2-1B-Instruct-q4f16_1-MLC",
+    },
+    {
+        id: "Llama-3.2-3B-Instruct-q4f16_1-MLC",
+        label: "Llama 3.2 3B Instruct q4f16 — stronger Meta fallback (~2.2 GB VRAM)",
+        shortName: "Llama 3.2 3B",
+        description: "More capable Llama option for machines that can handle a larger local browser model.",
+        approximateDownload: "large model download plus browser cache overhead",
+        vramRequiredMB: 2263.69,
+        vramSource: "webllm",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/Llama-3.2-3B-Instruct-q4f16_1-MLC",
+    },
+    {
+        id: "Hermes-3-Llama-3.2-3B-q4f16_1-MLC",
+        label: "Hermes 3 Llama 3.2 3B q4f16 — extraction/chat alternate (~2.2 GB VRAM)",
+        shortName: "Hermes 3 Llama 3.2 3B",
+        description: "Hermes-tuned Llama 3.2 3B variant. Useful for comparing structured-output behavior against the base Llama instruct build.",
+        approximateDownload: "large model download plus browser cache overhead",
+        vramRequiredMB: 2263.69,
+        vramSource: "webllm",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.06 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/Hermes-3-Llama-3.2-3B-q4f16_1-MLC",
+    },
+    {
+        id: "OLMo-2-0425-1B-Instruct-q4f16_1-MLC",
+        label: "OLMo 2 0425 1B Instruct q4f16 — open 1B alternate (~1.7 GB VRAM)",
+        shortName: "OLMo 2 1B",
+        description: "Recent small OLMo 2 instruction model compiled for WebLLM. Good for a non-Qwen, non-Llama comparison while staying near the 1B class.",
+        approximateDownload: "medium model download plus browser cache overhead",
+        vramRequiredMB: 1776.75,
+        vramSource: "webllm",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 512, topP: 0.9, repetitionPenalty: 1.06 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/OLMo-2-0425-1B-Instruct-q4f16_1-MLC",
+    },
+    {
+        id: "Phi-4-mini-instruct-q4f16_1-MLC",
+        label: "Phi-4 mini Instruct q4f16 — capable desktop option (~3.4 GB VRAM)",
+        shortName: "Phi-4 mini",
+        description: "Capable compact Phi model. Use on desktop-class devices when smaller models miss fields or produce lower-quality extraction.",
+        approximateDownload: "large model download plus browser cache overhead",
+        vramRequiredMB: 3437.58,
+        vramSource: "webllm",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/Phi-4-mini-instruct-q4f16_1-MLC",
+    },
+    {
+        id: "Ministral-3-3B-Instruct-2512-BF16-q4f16_1-MLC",
+        label: "Ministral 3 3B Instruct q4f16 — recent Mistral-family option (~2.8 GB VRAM)",
+        shortName: "Ministral 3 3B",
+        description: "Recent Mistral-family WebLLM option. Useful for comparing extraction quality against Qwen3.5/Llama/Phi on desktop-class GPUs.",
+        approximateDownload: "large model download plus browser cache overhead",
+        vramRequiredMB: 2863.69,
+        vramSource: "webllm",
+        lowResourceRequired: true,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/Ministral-3-3B-Instruct-2512-BF16-q4f16_1-MLC",
+    },
+    {
+        id: "Mistral-7B-Instruct-v0.3-q4f16_1-MLC",
+        label: "Mistral 7B Instruct v0.3 q4f16 — high-memory quality test (~4.5 GB VRAM)",
+        shortName: "Mistral 7B v0.3",
+        description: "Larger Mistral-family model. Keep this as a desktop/high-memory comparison model; it is not a first-load choice for casual users.",
+        approximateDownload: "very large model download plus browser cache overhead",
+        vramRequiredMB: 4573.39,
+        vramSource: "webllm",
+        lowResourceRequired: false,
+        disableThinkingSupported: false,
+        defaultRuntime: extractionPreset({ maxTokens: 768, topP: 0.9, repetitionPenalty: 1.05 }),
+        source: "webllm-prebuilt",
+        backend: "webllm",
+        stability: "stable",
+        officialRepo: "https://huggingface.co/mlc-ai/Mistral-7B-Instruct-v0.3-q4f16_1-MLC",
+    },
+];
+export function getModelPreset(modelId) {
+    return MODEL_PRESETS.find((preset) => preset.id === modelId) ?? MODEL_PRESETS[0];
+}
+export function formatMB(mb) {
+    if (!Number.isFinite(mb))
+        return "unknown";
+    return mb >= 1024 ? `${(mb / 1024).toFixed(2)} GB` : `${Math.round(mb)} MB`;
+}
+export function modelRepoLabel(preset) {
+    try {
+        const url = new URL(preset.artifactRepo ?? preset.officialRepo);
+        return url.pathname.replace(/^\//, "");
+    }
+    catch {
+        return preset.id;
+    }
+}
+//# sourceMappingURL=models.js.map