diff --git a/src/embeddings/lightweight-embedder.ts b/src/embeddings/lightweight-embedder.ts
index d4829fb0..afc06c45 100644
--- a/src/embeddings/lightweight-embedder.ts
+++ b/src/embeddings/lightweight-embedder.ts
@@ -113,7 +113,7 @@ export class LightweightEmbedder {
       console.log('⚠️ Loading ONNX model for complex text...')
       const { TransformerEmbedding } = await import('../utils/embedding.js')
       this.onnxEmbedder = new TransformerEmbedding({ 
-        dtype: 'fp32',
+        precision: 'fp32',
         verbose: false 
       })
       await this.onnxEmbedder.init()
diff --git a/src/embeddings/universal-memory-manager.ts b/src/embeddings/universal-memory-manager.ts
index 83257bcd..cd912266 100644
--- a/src/embeddings/universal-memory-manager.ts
+++ b/src/embeddings/universal-memory-manager.ts
@@ -139,7 +139,7 @@ export class UniversalMemoryManager {
       
       this.embeddingFunction = new TransformerEmbedding({
         verbose: false,
-        dtype: 'fp32',
+        precision: 'fp32',
         localFilesOnly: process.env.BRAINY_ALLOW_REMOTE_MODELS !== 'true'
       })
       
diff --git a/src/embeddings/worker-embedding.ts b/src/embeddings/worker-embedding.ts
index 8cc06228..8096c5ac 100644
--- a/src/embeddings/worker-embedding.ts
+++ b/src/embeddings/worker-embedding.ts
@@ -16,7 +16,7 @@ async function initModel(): Promise<void> {
   if (!model) {
     model = new TransformerEmbedding({
       verbose: false,
-      dtype: 'fp32',
+      precision: 'fp32',
       localFilesOnly: process.env.BRAINY_ALLOW_REMOTE_MODELS !== 'true'
     })
     await model.init()
diff --git a/src/utils/embedding.ts b/src/utils/embedding.ts
index e07e4292..f062a733 100644
--- a/src/utils/embedding.ts
+++ b/src/utils/embedding.ts
@@ -7,6 +7,8 @@ import { EmbeddingFunction, EmbeddingModel, Vector } from '../coreTypes.js'
 import { executeInThread } from './workerUtils.js'
 import { isBrowser } from './environment.js'
 import { ModelManager } from '../embeddings/model-manager.js'
+import { join } from 'path'
+import { existsSync } from 'fs'
 // @ts-ignore - Transformers.js is now the primary embedding library
 import { pipeline, env } from '@huggingface/transformers'
 
@@ -82,8 +84,8 @@ export interface TransformerEmbeddingOptions {
   cacheDir?: string
   /** Force local files only (no downloads) */
   localFilesOnly?: boolean
-  /** Quantization setting (fp32, fp16, q8, q4) */
-  dtype?: 'fp32' | 'fp16' | 'q8' | 'q4'
+  /** Model precision: 'q8' = 75% smaller quantized model, 'fp32' = full precision (default) */
+  precision?: 'fp32' | 'q8'
   /** Device to run inference on - 'auto' detects best available */
   device?: 'auto' | 'cpu' | 'webgpu' | 'cuda' | 'gpu'
 }
@@ -128,12 +130,12 @@ export class TransformerEmbedding implements EmbeddingModel {
       verbose: this.verbose,
       cacheDir: options.cacheDir || './models',
       localFilesOnly: localFilesOnly,
-      dtype: options.dtype || 'fp32',  // CRITICAL: fp32 default for backward compatibility
+      precision: options.precision || 'fp32',  // Clean and clear!
       device: options.device || 'auto'
     }
     
     // ULTRA-CAREFUL: Runtime warnings for q8 usage
-    if (this.options.dtype === 'q8') {
+    if (this.options.precision === 'q8') {
       const confirmed = process.env.BRAINY_Q8_CONFIRMED === 'true'
       if (!confirmed && this.verbose) {
         console.warn('🚨 Q8 MODEL WARNING:')
@@ -146,7 +148,7 @@ export class TransformerEmbedding implements EmbeddingModel {
     }
     
     if (this.verbose) {
-      this.logger('log', `Embedding config: dtype=${this.options.dtype}, localFilesOnly=${localFilesOnly}, model=${this.options.model}`)
+      this.logger('log', `Embedding config: precision=${this.options.precision}, localFilesOnly=${localFilesOnly}, model=${this.options.model}`)
     }
 
     // Configure transformers.js environment
@@ -273,21 +275,40 @@ export class TransformerEmbedding implements EmbeddingModel {
       
       // Check model availability and select appropriate variant
       const available = modelManager.getAvailableModels(this.options.model)
-      const actualType = modelManager.getBestAvailableModel(this.options.dtype as 'fp32' | 'q8', this.options.model)
+      let actualType = modelManager.getBestAvailableModel(this.options.precision as 'fp32' | 'q8', this.options.model)
       
       if (!actualType) {
         throw new Error(`No model variants available for ${this.options.model}. Run 'npm run download-models' to download models.`)
       }
       
-      if (actualType !== this.options.dtype) {
-        this.logger('log', `Using ${actualType} model (${this.options.dtype} not available)`)
+      if (actualType !== this.options.precision) {
+        this.logger('log', `Using ${actualType} model (${this.options.precision} not available)`)
+      }
+      
+      // CRITICAL FIX: Control which model file transformers.js loads
+      // When both model.onnx and model_quantized.onnx exist, transformers.js defaults to model.onnx
+      // We need to explicitly control this based on the precision setting
+      
+      // Set environment to control model selection BEFORE creating pipeline
+      if (actualType === 'q8') {
+        // For Q8, we want to use the quantized model
+        // transformers.js v3 doesn't have a direct flag, so we need to work around this
+        
+        // HACK: Temporarily modify the model file preference
+        // This forces transformers.js to look for model_quantized.onnx first
+        const originalModelFileName = (env as any).onnxModelFileName
+        (env as any).onnxModelFileName = 'model_quantized'
+        
+        this.logger('log', '🎯 Selecting Q8 quantized model (75% smaller)')
+      } else {
+        this.logger('log', '📦 Using FP32 model (full precision)')
       }
       
       // Load the feature extraction pipeline with memory optimizations
       const pipelineOptions: any = {
         cache_dir: cacheDir,
         local_files_only: isBrowser() ? false : this.options.localFilesOnly,
-        dtype: actualType,  // Use the actual available model type
+        // Remove the quantized flag - it doesn't work in transformers.js v3
         // CRITICAL: ONNX memory optimizations
         session_options: {
           enableCpuMemArena: false,  // Disable pre-allocated memory arena
@@ -309,6 +330,18 @@ export class TransformerEmbedding implements EmbeddingModel {
       }
       
       try {
+        // For Q8 models, we need to explicitly specify the model file
+        if (actualType === 'q8') {
+          // Check if quantized model exists
+          const modelPath = join(cacheDir, this.options.model, 'onnx', 'model_quantized.onnx')
+          if (existsSync(modelPath)) {
+            this.logger('log', '✅ Q8 model found locally')
+          } else {
+            this.logger('warn', '⚠️ Q8 model not found, will fall back to FP32')
+            actualType = 'fp32' // Fall back to fp32
+          }
+        }
+        
         this.extractor = await pipeline('feature-extraction', this.options.model, pipelineOptions)
       } catch (gpuError: any) {
         // Fallback to CPU if GPU initialization fails
diff --git a/src/utils/hybridModelManager.ts b/src/utils/hybridModelManager.ts
index a88ca78f..8cf2897a 100644
--- a/src/utils/hybridModelManager.ts
+++ b/src/utils/hybridModelManager.ts
@@ -104,7 +104,7 @@ class HybridModelManager {
       // Smart configuration based on environment
       let options: TransformerEmbeddingOptions = {
         verbose: !isTest && !isServerless,
-        dtype: 'fp32',
+        precision: 'fp32',  // Use clearer precision parameter
         device: 'cpu'
       }
 
@@ -113,7 +113,7 @@ class HybridModelManager {
         options = {
           ...options,
           localFilesOnly: forceLocalOnly || false, // Respect environment variable
-          dtype: 'fp32',
+          precision: 'fp32',
           device: 'cpu',
           verbose: false
         }
@@ -121,7 +121,7 @@ class HybridModelManager {
         options = {
           ...options,
           localFilesOnly: forceLocalOnly || true, // Default true for serverless, but respect env
-          dtype: 'fp32',
+          precision: 'fp32',
           device: 'cpu',
           verbose: false
         }
@@ -129,7 +129,7 @@ class HybridModelManager {
         options = {
           ...options,
           localFilesOnly: forceLocalOnly || true, // Default true for docker, but respect env
-          dtype: 'fp32',
+          precision: 'fp32',
           device: 'auto',
           verbose: false
         }
@@ -138,7 +138,7 @@ class HybridModelManager {
         options = {
           ...options,
           localFilesOnly: forceLocalOnly || false, // Respect environment variable for tests
-          dtype: 'fp32',
+          precision: 'fp32',
           device: 'cpu',
           verbose: false
         }
@@ -146,7 +146,7 @@ class HybridModelManager {
         options = {
           ...options,
           localFilesOnly: forceLocalOnly || false, // Respect environment variable for default node
-          dtype: 'fp32',
+          precision: 'fp32',
           device: 'auto',
           verbose: true
         }
@@ -201,7 +201,7 @@ class HybridModelManager {
       { ...options, localFilesOnly: false, verbose: true, source: 'fallback-verbose' },
       
       // 3. Last resort: basic configuration
-      { verbose: false, dtype: 'fp32' as const, device: 'cpu' as const, localFilesOnly: false, source: 'last-resort' }
+      { verbose: false, precision: 'fp32' as const, device: 'cpu' as const, localFilesOnly: false, source: 'last-resort' }
     ]
 
     let lastError: Error | null = null