@missionsquad/browserai 0.1.0 → 0.2.0

package/src/client.ts

@@ -1,6 +1,6 @@
 import type { MLCEngineInterface } from "@mlc-ai/web-llm";
-import type { ModelPreset, ModelTask } from "./models.js";
-import { getModelPreset, MODEL_PRESETS, modelRepoLabel, modelSlots } from "./models.js";
+import type { ModelPreset, ModelTask, ModelBackend } from "./models.js";
+import { getModelPreset, MODEL_PRESETS, modelRepoLabel, modelSlots, supportsWasmFallback } from "./models.js";
 import type {
   ChatMessage,
   EngineHandles,
@@ -13,7 +13,6 @@ import type {
 import { EMPTY_RESPONSE, generate } from "./inference.js";
 import type { RuntimeParameters } from "./generation.js";
 import { clampNumber, runtimeParamsFromPreset } from "./generation.js";
-import type { MultimodalRuntime } from "./models.js";
 import type {
   MultimodalDevice,
   MultimodalHandle,
@@ -22,9 +21,11 @@ import type {
   SttResult,
   TtsTextSplitter,
   VlmResult,
+  ImageClassificationResult,
 } from "./multimodal.js";
 import {
   describeImage as describeImageWithHandle,
+  classifyImage as classifyImageWithHandle,
   disposeMultimodal,
   generateGemmaText,
   loadMultimodal,
@@ -39,7 +40,7 @@ import type { ModelSourceMode } from "./model-source.js";
 import { applyContextOverride, buildWebLLMAppConfig, TRANSFORMERS_CACHE_KEY } from "./model-source.js";
 import { verifyTransformersProxy, verifyWebLLMProxy } from "./proxy-verify.js";
 import type { HardwareSnapshot } from "./hardware.js";
-import { collectHardwareSnapshot } from "./hardware.js";
+import { collectHardwareSnapshot, isChromiumBased } from "./hardware.js";
 import type { CacheCleanupResult, CachedModelStatus, IndividualModelCleanupResult, ModelCacheTarget, StorageEstimateSnapshot } from "./model-cache.js";
 import {
   deleteOneModelArtifactsFromBrowserStorage,
@@ -65,7 +66,7 @@ import { synthesizeClip } from "./audio/clip.js";
 export type LoadedModel = {
   modelId: string;
   slots: ModelTask[];
-  backend: "webllm" | "transformers-js";
+  backend: ModelBackend;
   engine: MLCEngineInterface | null;
   worker: Worker | null;
   transformersGenerator: TransformersTextGenerationPipeline | null;
@@ -81,7 +82,6 @@ export type LoadProgress = { modelId: string; progress: number; status?: string;
  * Multimodal runtimes whose loaders honor a wasm device. The VLM/Gemma/audio-LLM recipes hardcode
  * WebGPU (their verified recipes only run there), so they cannot fall back.
  */
-const WASM_CAPABLE_RUNTIMES: ReadonlySet<MultimodalRuntime> = new Set(["stt", "tts-pipeline", "tts-kokoro"]);
 
 export type BrowserAIEvents = {
   /** Download/initialization progress for the model currently loading (progress 0–1). */
@@ -261,11 +261,16 @@ export class BrowserAI extends TypedEmitter<BrowserAIEvents> {
     try {
       this.#hardware = await collectHardwareSnapshot();
       let device: MultimodalDevice = "webgpu";
+      const wasmCapable = supportsWasmFallback(preset);
       if (!this.#hardware.webgpuSupported) {
         // Only the runtimes whose loaders honor a wasm device (ASR pipelines, TTS) can fall back;
         // VLM/Gemma/audio-LLM recipes hardcode WebGPU, and text engines require it outright.
-        if (preset.mmRuntime && WASM_CAPABLE_RUNTIMES.has(preset.mmRuntime)) device = "wasm";
+        if (wasmCapable) device = "wasm";
         else throw new WebGPUUnavailableError(this.#hardware.webgpuReason);
+      } else if (wasmCapable && !isChromiumBased()) {
+        // Firefox and Safari now pass the WebGPU probe, but onnxruntime-web's WebGPU backend only
+        // reliably runs on Chromium (shader miscompiles / quantized-session failures elsewhere).
+        device = "wasm";
       }
 
       // Free any loaded models occupying the slots this one needs, waiting out their in-flight runs.
@@ -437,6 +442,18 @@ export class BrowserAI extends TypedEmitter<BrowserAIEvents> {
   }
 
   async #loadMultimodalModel(preset: ModelPreset, device: MultimodalDevice, progress: (info: Omit<LoadProgress, "modelId">) => void): Promise<LoadedModel> {
+    if (preset.mmRuntime === "stt-tdt") {
+      this.#status(`Loading ${preset.shortName} with ONNX Runtime…`);
+      const root = this.#modelSourceMode() === "same-origin-proxy"
+        ? `${this.#requireProxyOrigin()}/hf-transformers/` : "https://huggingface.co/";
+      if (this.#modelSourceMode() === "same-origin-proxy" && this.config.verifyProxy) {
+        await verifyTransformersProxy(this.#requireProxyOrigin(), preset.id);
+      }
+      const { loadTdtModel } = await import("./tdt/index.js");
+      const model = await loadTdtModel(preset.id, { ...this.config.tdt, device, baseUrl: `${root}${preset.id}/resolve/main/`, onProgress: this.#makeTransformersProgressForwarder(progress) });
+      progress({ progress: 1 });
+      return { ...newLoadedModel(preset), multimodal: { kind: "tdt", model } };
+    }
     this.#status(`Loading ${preset.shortName} with Transformers.js/WebGPU…`);
     progress({ progress: 0.05 });
 
@@ -542,12 +559,12 @@ export class BrowserAI extends TypedEmitter<BrowserAIEvents> {
     });
   }
 
-  /** Transcribe 16 kHz mono PCM (or a URL) with the loaded Transcription-slot model. */
-  async transcribe(audio: Float32Array | string, onDelta?: (text: string) => void): Promise<SttResult> {
+  /** Transcribe 16 kHz mono PCM. ASR pipelines also accept URLs; TDT and generative models require decoded PCM. The optional signal cooperatively cancels TDT decoding between inference calls. */
+  async transcribe(audio: Float32Array | string, onDelta?: (text: string) => void, signal?: AbortSignal): Promise<SttResult> {
     const model = this.sttModel();
     if (!model?.multimodal) throw new MissingModelError("stt");
     const handle = model.multimodal;
-    return this.#runExclusive(model, () => transcribeWithHandle(handle, audio, onDelta));
+    return this.#runExclusive(model, () => transcribeWithHandle(handle, audio, onDelta, signal));
   }
 
   /** Synthesize speech with the loaded Speech-slot model. Returns the raw model audio. */
@@ -620,6 +637,14 @@ export class BrowserAI extends TypedEmitter<BrowserAIEvents> {
     };
   }
 
+  /** Return scored labels with the loaded Image-slot classifier. */
+  async classifyImage(imageSrc: string): Promise<ImageClassificationResult> {
+    const model = this.visionModel();
+    if (!model?.multimodal) throw new MissingModelError("vision");
+    const handle = model.multimodal;
+    return this.#runExclusive(model, () => classifyImageWithHandle(handle, imageSrc));
+  }
+
   /** Ask a question about an image with the loaded Image-slot model (VLM or Gemma). */
   async describeImage(imageSrc: string, prompt: string, onDelta?: (text: string) => void): Promise<VlmResult> {
     const model = this.visionModel();

package/src/config.ts

@@ -26,6 +26,9 @@ export type KokoroOptions = {
   load?: () => Promise<KokoroModule>;
 };
 
+/** Optional location of the matching ORT 1.31.1 asyncify MJS/WASM files. */
+export type TdtOptions = { wasmPaths?: string | { mjs: string; wasm: string } };
+
 export type TtsDefaults = {
   /** Voice id for the loaded TTS model's roster (Kokoro: "af_heart" etc.; Supertonic: "F1"–"M5"). */
   voice: string;
@@ -69,6 +72,8 @@ export type BrowserAIConfig = {
   cacheBackend?: BrowserCacheBackend;
   webllm?: WebLLMOptions;
   kokoro?: KokoroOptions;
+  /** Override speech-runtime assets for self-hosted/offline applications. */
+  tdt?: TdtOptions;
   /** Default TTS voice/speed/pitch used by synthesize helpers and the voice pipeline. */
   tts?: Partial<TtsDefaults>;
   history?: HistoryOptions;
@@ -81,6 +86,7 @@ export type ResolvedConfig = {
   cacheBackend: BrowserCacheBackend;
   webllm: WebLLMOptions;
   kokoro: KokoroOptions;
+  tdt: TdtOptions;
   tts: TtsDefaults;
   history: HistoryOptions;
 };
@@ -98,6 +104,7 @@ export function resolveConfig(config: BrowserAIConfig = {}): ResolvedConfig {
     cacheBackend: config.cacheBackend ?? "indexeddb",
     webllm: config.webllm ?? {},
     kokoro: config.kokoro ?? {},
+    tdt: config.tdt ?? {},
     tts: { voice: "af_heart", speed: 1, pitch: 1, ...config.tts },
     history: config.history ?? {},
   };

package/src/hardware.ts

@@ -19,6 +19,21 @@ export type HardwareSnapshot = {
   platform?: string;
 };
 
+/**
+ * Whether this is a Chromium-based browser (Chrome, Edge, Opera, Brave, …), per the UA-Client-Hints
+ * brand list. onnxruntime-web's WebGPU backend is developed and tested against Chromium; Firefox's
+ * and Safari's newer WebGPU implementations still fail on some of its generated WGSL and its
+ * quantized-weight session transforms (errors that only surface at load/inference time), so
+ * wasm-capable runtimes prefer wasm off-Chromium even when the WebGPU probe passes.
+ * navigator.userAgentData is itself Chromium-only, and every Chromium new enough for WebGPU (113+)
+ * ships it, so the brand list is a reliable signal.
+ */
+export function isChromiumBased(): boolean {
+  if (typeof navigator === "undefined") return false;
+  const nav = navigator as Navigator & { userAgentData?: { brands?: Array<{ brand: string }> } };
+  return nav.userAgentData?.brands?.some((entry) => entry.brand === "Chromium") ?? false;
+}
+
 const LIMIT_NAMES = [
   "maxBufferSize",
   "maxStorageBufferBindingSize",

package/src/history.ts

@@ -36,7 +36,7 @@ export type RunRecord = {
   /** Derived vendor bucket for filtering/labels. */
   vendor: ModelVendor;
   /** Backend used for the run. */
-  backend: "webllm" | "transformers-js";
+  backend: "webllm" | "transformers-js" | "onnx-runtime";
   /** Task key ({@link RunCaseKey} suggested; any non-empty string accepted). */
   case: string;
   promptTokens: number;
@@ -111,7 +111,7 @@ function coerceRecord(value: unknown): RunRecord | null {
     ts: record.ts as number,
     model: record.model,
     vendor: VALID_VENDORS.has(record.vendor as ModelVendor) ? (record.vendor as ModelVendor) : "other",
-    backend: record.backend === "transformers-js" ? "transformers-js" : "webllm",
+    backend: record.backend === "onnx-runtime" ? "onnx-runtime" : record.backend === "transformers-js" ? "transformers-js" : "webllm",
     case: record.case,
     promptTokens: record.promptTokens as number,
     completionTokens: record.completionTokens as number,

package/src/index.ts

@@ -12,7 +12,7 @@ export type { BrowserAIEvents, GenerateTextOptions, LoadOptions, LoadProgress, L
 
 // Configuration
 export { DEFAULT_CONTEXT_LENGTH, resolveConfig } from "./config.js";
-export type { BrowserAIConfig, HistoryOptions, KokoroOptions, ResolvedConfig, StorageLike, TtsDefaults, WebLLMOptions } from "./config.js";
+export type { BrowserAIConfig, HistoryOptions, KokoroOptions, ResolvedConfig, StorageLike, TdtOptions, TtsDefaults, WebLLMOptions } from "./config.js";
 
 // Errors
 export {
@@ -81,6 +81,7 @@ export type {
 // Multimodal runtimes (STT / TTS / VLM / Gemma)
 export {
   capSrc,
+  classifyImage,
   DEFAULT_KOKORO_CDN,
   describeImage,
   disposeMultimodal,
@@ -99,6 +100,7 @@ export {
 export type {
   AudioWindow,
   GenStats,
+  ImageClassificationResult,
   KokoroModule,
   LoadMultimodalOptions,
   MultimodalDevice,

package/src/inference.ts

@@ -305,9 +305,10 @@ async function runTransformers(
 ): Promise<TransformersRunResult> {
   const generator = handles.transformersGenerator!;
   const tokenizer = handles.transformersTokenizer ?? generator.tokenizer ?? null;
+  const chatTemplateOptions = preset.disableThinkingSupported ? { enable_thinking: !runtime.disableThinking } : undefined;
 
   const telemetry: TransformersGenerationTelemetry = {
-    promptTokens: countPromptTokens(tokenizer, messages),
+    promptTokens: countPromptTokens(tokenizer, messages, chatTemplateOptions),
     tokenCallbackIntervalSeconds: [],
     tokenCallbackCount: 0,
     tokenCountSource: "unavailable",
@@ -323,8 +324,8 @@ async function runTransformers(
   };
   if (runtime.temperature > 0) generationOptions.temperature = runtime.temperature;
   if (runtime.seed !== null) generationOptions.seed = runtime.seed;
-  if (preset.disableThinkingSupported) {
-    generationOptions.tokenizer_encode_kwargs = { enable_thinking: !runtime.disableThinking };
+  if (chatTemplateOptions) {
+    generationOptions.tokenizer_encode_kwargs = chatTemplateOptions;
   }
 
   const Streamer = handles.transformersModule?.TextStreamer;
@@ -438,7 +439,7 @@ export function collectGenericStats(model: string, backend: ModelBackend | "unkn
 /* ------------------------------------------------------------------ */
 
 export function buildRuntimeStatsText(stats: RuntimeStats): string {
-  const backendLabel = stats.backend === "transformers-js" ? "Transformers.js" : stats.backend === "webllm" ? "WebLLM" : "unknown";
+  const backendLabel = stats.backend === "onnx-runtime" ? "ONNX Runtime" : stats.backend === "transformers-js" ? "Transformers.js" : stats.backend === "webllm" ? "WebLLM" : "unknown";
   const lines = [
     `model: ${stats.model}`,
     `backend: ${backendLabel}`,
@@ -514,11 +515,16 @@ function extractGeneratedText(output: unknown): string {
   return JSON.stringify(output, null, 2);
 }
 
-function countPromptTokens(tokenizer: TransformersTokenizer | null, messages: ChatMessage[]): number | undefined {
+function countPromptTokens(
+  tokenizer: TransformersTokenizer | null,
+  messages: ChatMessage[],
+  chatTemplateOptions?: Record<string, unknown>,
+): number | undefined {
   if (!tokenizer) return undefined;
   try {
     if (typeof tokenizer.apply_chat_template === "function") {
-      const tokenized = tokenizer.apply_chat_template(messages, { tokenize: true, add_generation_prompt: true });
+      // Count the same prompt the pipeline generates, including MiniCPM5's optional empty think block.
+      const tokenized = tokenizer.apply_chat_template(messages, { tokenize: true, add_generation_prompt: true, ...chatTemplateOptions });
       const count = extractTokenCount(tokenized);
       if (typeof count === "number") return count;
     }

package/src/metrics.ts

@@ -14,7 +14,7 @@ export function buildMultimodalStats(gen: GenStats, modelId: string): RuntimeSta
   return {
     rawText: "",
     model: modelId,
-    backend: "transformers-js",
+    backend: getModelPreset(modelId).backend,
     finishReason: "stop",
     measuredElapsedMs: elapsedMs,
     completionTokens,
@@ -41,7 +41,7 @@ export function runRecordFromStats(caseKey: string, stats: RuntimeStats): RunRec
     ts: Date.now(),
     model: preset.shortName,
     vendor: vendorForModel(preset),
-    backend: stats.backend === "transformers-js" ? "transformers-js" : "webllm",
+    backend: stats.backend === "onnx-runtime" ? "onnx-runtime" : stats.backend === "transformers-js" ? "transformers-js" : "webllm",
     case: caseKey,
     promptTokens: stats.promptTokens ?? 0,
     completionTokens: stats.completionTokens ?? 0,

package/src/models.ts

@@ -11,7 +11,7 @@ export type RuntimePreset = {
   disableThinking: boolean;
 };
 
-export type ModelBackend = "webllm" | "transformers-js";
+export type ModelBackend = "webllm" | "transformers-js" | "onnx-runtime";
 
 /** Input/output modality of a model, used for the library modality-flow line + Playground gating. */
 export type Modality = "text" | "image" | "audio";
@@ -23,9 +23,11 @@ export type ModelTask = "text" | "vision" | "stt" | "tts";
  * inference recipe in llm/multimodal.ts.
  */
 export type MultimodalRuntime =
+  | "stt-tdt" // ONNX Runtime TDT speech recognition
   | "stt" // automatic-speech-recognition pipeline (Whisper / Moonshine)
   | "tts-kokoro" // kokoro-js KokoroTTS
   | "tts-pipeline" // native text-to-speech pipeline (Supertonic)
+  | "image-classification" // scored image labels (NSFW detection)
   | "vlm-vision2seq" // AutoModelForVision2Seq (SmolVLM)
   | "vlm-imagetext" // AutoModelForImageTextToText (Qwen3-VL, GLM-OCR)
   | "audio-text-to-text" // AutoModelForAudioTextToText with an <audio> token in the prompt (Granite Speech, Ultravox, non-realtime Voxtral)
@@ -60,6 +62,8 @@ export type ModelPreset = {
   modalityOut?: Modality[];
   /** How a multimodal (non-LLM-text) model loads + runs. Undefined for plain text LLMs. */
   mmRuntime?: MultimodalRuntime;
+  /** Per-model image processing and generation behavior. */
+  visionOptions?: { defaultPrompt?: string; maxNewTokens?: number; preserveSpecialTokens?: boolean; doImageSplitting?: boolean; outputFormat?: "doctags" };
   /** Optional display brand for the vendor tag (e.g. "Moonshine", "Kokoro", "Zhipu"). Falls back to the vendor label. */
   brand?: string;
 };
@@ -322,6 +326,80 @@ export const MODEL_PRESETS: ModelPreset[] = [
     officialRepo: "https://huggingface.co/google/gemma-4-E4B-it",
     artifactRepo: "https://huggingface.co/welcoma/gemma-4-E4B-it-q4f16_1-MLC",
   },
+  {
+    id: "enriqueag/phonon-2.5-onnx",
+    label: "Phonon 2.5 — compact speech recognition via WebGPU + WASM",
+    shortName: "Phonon 2.5",
+    description: "Fermion speech recognition with a compact quantized WebGPU encoder and a WASM token-duration decoder. Audio stays on device.",
+    approximateDownload: "~381 MB encoder, decoder and vocabulary",
+    vramRequiredMB: 1536,
+    vramSource: "estimated",
+    lowResourceRequired: true,
+    disableThinkingSupported: false,
+    defaultRuntime: extractionPreset(),
+    source: "transformers-onnx",
+    backend: "onnx-runtime",
+    stability: "experimental",
+    officialRepo: "https://huggingface.co/FermionResearch/Phonon-2.5",
+    artifactRepo: "https://huggingface.co/enriqueag/phonon-2.5-onnx",
+    tasks: ["stt"], modalityIn: ["audio"], modalityOut: ["text"],
+    mmRuntime: "stt-tdt",
+    brand: "Fermion",
+    notes: "Requires WebGPU for the compact encoder; the decoder runs on CPU/WASM. Long audio is split near quiet intervals into chunks of at most 30 seconds. Memory is estimated.",
+  },
+  {
+    id: "istupakov/parakeet-tdt-0.6b-v3-onnx",
+    label: "Parakeet TDT 0.6B v3 — multilingual speech recognition via WASM",
+    shortName: "Parakeet TDT v3",
+    description: "NVIDIA multilingual token-duration speech recognition, using the published int8 ONNX encoder and decoder on CPU/WASM.",
+    approximateDownload: "~671 MB int8 encoder, decoder, frontend and vocabulary",
+    vramRequiredMB: 0,
+    vramSource: "estimated",
+    lowResourceRequired: false,
+    disableThinkingSupported: false,
+    defaultRuntime: extractionPreset(),
+    source: "transformers-onnx",
+    backend: "onnx-runtime",
+    stability: "experimental",
+    officialRepo: "https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3",
+    artifactRepo: "https://huggingface.co/istupakov/parakeet-tdt-0.6b-v3-onnx",
+    tasks: ["stt"], modalityIn: ["audio"], modalityOut: ["text"],
+    mmRuntime: "stt-tdt",
+    brand: "NVIDIA",
+    notes: "Runs on CPU/WASM and consumes system memory despite requiring no GPU memory. Long audio is split near quiet intervals into chunks of at most 30 seconds.",
+  },
+  {
+    id: "onnx-community/granite-docling-258M-ONNX",
+    label: "Granite Docling 258M — document conversion to DocTags",
+    shortName: "Granite Docling 258M",
+    description: "IBM Granite document vision model. Converts page images into structured DocTags, preserving layout and document markup.",
+    approximateDownload: "~1.27 GB FP32 ONNX modules plus tokenizer/config files",
+    vramRequiredMB: 3072, vramSource: "estimated", lowResourceRequired: false,
+    disableThinkingSupported: false, defaultRuntime: extractionPreset({ maxTokens: 4096 }),
+    source: "transformers-onnx", backend: "transformers-js", stability: "experimental",
+    officialRepo: "https://huggingface.co/ibm-granite/granite-docling-258M",
+    artifactRepo: "https://huggingface.co/onnx-community/granite-docling-258M-ONNX",
+    transformersDtype: "fp32", tasks: ["vision"], modalityIn: ["image", "text"], modalityOut: ["text"],
+    mmRuntime: "vlm-vision2seq", brand: "IBM",
+    visionOptions: { defaultPrompt: "Convert this page to docling.", maxNewTokens: 4096, preserveSpecialTokens: true, doImageSplitting: true, outputFormat: "doctags" },
+    notes: "FP32 is intentional: the FP16/q4f16 WebGPU paths failed accuracy smoke checks on Apple GPU. FP32 produced the expected result.",
+  },
+  {
+    id: "onnx-community/nsfw_image_detection-ONNX",
+    label: "NSFW image detection — scored normal / nsfw classification",
+    shortName: "NSFW image detection",
+    description: "ViT image classifier returning normal and nsfw scores. Scores are model predictions, not a guarantee of content safety.",
+    approximateDownload: "~343 MB FP32 ONNX weights plus image processor/config files",
+    vramRequiredMB: 768, vramSource: "estimated", lowResourceRequired: true,
+    disableThinkingSupported: false, defaultRuntime: extractionPreset({}),
+    source: "transformers-onnx", backend: "transformers-js", stability: "experimental",
+    officialRepo: "https://huggingface.co/Falconsai/nsfw_image_detection",
+    artifactRepo: "https://huggingface.co/onnx-community/nsfw_image_detection-ONNX",
+    transformersDtype: "fp32", tasks: ["vision"], modalityIn: ["image"], modalityOut: ["text"],
+    mmRuntime: "image-classification", brand: "Falconsai",
+    notes: "FP32 is intentional: the FP16/q4f16 WebGPU paths failed accuracy smoke checks on Apple GPU. FP32 produced the expected result.",
+  },
+
   {
     id: "onnx-community/whisper-base",
     label: "Whisper Base — multilingual speech-to-text via Transformers.js/WebGPU",
@@ -948,6 +1026,48 @@ export const MODEL_PRESETS: ModelPreset[] = [
     artifactRepo: "https://huggingface.co/onnx-community/granite-4.0-h-micro-ONNX",
     transformersDtype: "q4f16",
   },
+  {
+    id: "RASMUS/MiniCPM5-2B-ONNX",
+    label: "MiniCPM5 2B q4f16 — Transformers.js/WebGPU (~3 GB VRAM est.)",
+    shortName: "MiniCPM5 2B",
+    description:
+      "OpenBMB's regular MiniCPM5 2B instruction model, quantized for browser GPU inference with streaming, optional thinking and measured token/timing metrics.",
+    approximateDownload: "~1.83 GB ONNX weights plus tokenizer and browser cache overhead",
+    vramRequiredMB: 3072,
+    vramSource: "estimated",
+    lowResourceRequired: false,
+    disableThinkingSupported: true,
+    defaultRuntime: extractionPreset({ maxTokens: 768, disableThinking: true }),
+    source: "transformers-onnx",
+    backend: "transformers-js",
+    stability: "experimental",
+    officialRepo: "https://huggingface.co/openbmb/MiniCPM5-2B",
+    artifactRepo: "https://huggingface.co/RASMUS/MiniCPM5-2B-ONNX",
+    transformersDtype: "q4f16",
+    brand: "OpenBMB",
+    notes: "Requires WebGPU shader-f16. JSON schema mode is prompt-only. Browser memory depends on context length; the VRAM figure is an estimate.",
+  },
+  {
+    id: "j4ys0n/MiniCPM5-2B-heretic-abliterated-ONNX",
+    label: "MiniCPM5 2B Heretic q4f16 — Transformers.js/WebGPU (~3 GB VRAM est.)",
+    shortName: "MiniCPM5 2B Heretic",
+    description:
+      "Heretic-abliterated MiniCPM5 2B, converted from the original insraq safetensors checkpoint to browser-ready ONNX with streaming, optional thinking and measured token/timing metrics.",
+    approximateDownload: "~1.83 GB ONNX weights plus tokenizer and browser cache overhead",
+    vramRequiredMB: 3072,
+    vramSource: "estimated",
+    lowResourceRequired: false,
+    disableThinkingSupported: true,
+    defaultRuntime: extractionPreset({ maxTokens: 768, disableThinking: true }),
+    source: "transformers-onnx",
+    backend: "transformers-js",
+    stability: "experimental",
+    officialRepo: "https://huggingface.co/insraq/MiniCPM5-2B-heretic-abliterated",
+    artifactRepo: "https://huggingface.co/j4ys0n/MiniCPM5-2B-heretic-abliterated-ONNX",
+    transformersDtype: "q4f16",
+    brand: "OpenBMB",
+    notes: "Original Heretic checkpoint referenced by Abiray's model card. Requires WebGPU shader-f16. JSON schema mode is prompt-only; VRAM usage is estimated.",
+  },
   {
     id: "Mike0021/MiniCPM5-1B-ONNX-Web",
     label: "MiniCPM5 1B ONNX-web — community q4 Transformers.js/WebGPU export",
@@ -986,6 +1106,31 @@ export const MODEL_PRESETS: ModelPreset[] = [
     artifactRepo: "https://huggingface.co/HuggingFaceTB/SmolLM3-3B-ONNX",
     transformersDtype: "q4",
   },
+  {
+    id: "j4ys0n/NVIDIA-Nemotron-3.5-Bolt-E4B-ONNX",
+    label: "NVIDIA Nemotron 3.5 Bolt E4B ONNX — text, image and audio input; text output (~5.21 GB)",
+    shortName: "Nemotron 3.5 Bolt E4B",
+    description:
+      "NVIDIA's post-trained Gemma 4 E4B, exported for Transformers.js/WebGPU with text, image, and audio input and text output.",
+    approximateDownload: "~5.21 GB q4f16 ONNX artifacts plus tokenizer/config files",
+    vramRequiredMB: 8192,
+    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/nvidia/NVIDIA-Nemotron-3.5-Bolt-E4B-BF16",
+    artifactRepo: "https://huggingface.co/j4ys0n/NVIDIA-Nemotron-3.5-Bolt-E4B-ONNX",
+    notes: "Text, image, and audio browser inference smoke-tested on Apple WebGPU. Output quality and GPU memory have not been benchmarked; GPU memory is an estimate, not a measured minimum.",
+    transformersDtype: "q4f16",
+    tasks: ["text", "vision", "stt"],
+    modalityIn: ["text", "image", "audio"],
+    modalityOut: ["text"],
+    mmRuntime: "gemma-mm",
+    brand: "NVIDIA",
+  },
   {
     id: "onnx-community/gemma-4-E2B-it-ONNX",
     label: "Gemma 4 E2B ONNX — experimental Transformers.js/WebGPU alternative",
@@ -1169,3 +1314,9 @@ export function modelRepoLabel(preset: ModelPreset): string {
     return preset.id;
   }
 }
+
+/** Whether this preset has an implemented CPU/WASM load path. */
+export function supportsWasmFallback(preset: ModelPreset): boolean {
+  return preset.id === "istupakov/parakeet-tdt-0.6b-v3-onnx" ||
+    (preset.mmRuntime !== undefined && ["stt", "tts-pipeline", "tts-kokoro", "image-classification"].includes(preset.mmRuntime));
+}