diff --git a/Sources/FluidAudio/ModelNames.swift b/Sources/FluidAudio/ModelNames.swift
index edfa91bd..42d82612 100644
--- a/Sources/FluidAudio/ModelNames.swift
+++ b/Sources/FluidAudio/ModelNames.swift
@@ -317,12 +317,21 @@ public enum Repo: String, CaseIterable, Sendable {
 /// Encoder precision for the v3 Parakeet TDT 0.6B encoder.
 public enum ParakeetEncoderPrecision: String, Sendable, CaseIterable {
     case int8
+    /// Opt-in int8 per-channel linear re-quantization of the v3 encoder
+    /// (`Encoder_v2.mlmodelc`, 568M vs 425M). Avoids token corruption the
+    /// original 6-bit-LUT palettized `Encoder.mlmodelc` exhibits under
+    /// specific right-context (issue #760). `.int8` remains the default and
+    /// keeps loading the original file; select this explicitly to use the
+    /// rebuild.
+    case int8V2 = "int8-v2"
     case int4
 
     public var encoderFileName: String {
         switch self {
         case .int8:
             return ModelNames.ASR.encoderFile
+        case .int8V2:
+            return ModelNames.ASR.encoderV2File
         case .int4:
             return ModelNames.ASR.encoderInt4File
         }
@@ -391,6 +400,10 @@ public enum ModelNames {
         /// Joint decoder variant for v3 that exposes top-K outputs
         /// (`top_k_ids`, `top_k_logits`) used for language-aware script filtering.
         public static let jointV3File = "JointDecisionv3.mlmodelc"
+        /// v3 encoder re-quantized as int8 per-channel linear (issue #760).
+        /// Published alongside the immutable original `Encoder.mlmodelc`;
+        /// HF repo files are never mutated in place, fixes ship as new names.
+        public static let encoderV2File = "Encoder_v2.mlmodelc"
         public static let encoderInt4File = "EncoderInt4.mlmodelc"
         public static let ctcHeadFile = ctcHead + ".mlmodelc"
 
diff --git a/Sources/FluidAudioCLI/Commands/ASR/Parakeet/SlidingWindow/TranscribeCommand.swift b/Sources/FluidAudioCLI/Commands/ASR/Parakeet/SlidingWindow/TranscribeCommand.swift
index 838396da..6e3844bf 100644
--- a/Sources/FluidAudioCLI/Commands/ASR/Parakeet/SlidingWindow/TranscribeCommand.swift
+++ b/Sources/FluidAudioCLI/Commands/ASR/Parakeet/SlidingWindow/TranscribeCommand.swift
@@ -1023,7 +1023,8 @@ enum TranscribeCommand {
                 --output-json <file>           Save full transcription to JSON
                 --model-version <v2|v3|110m>   ASR model version (default: v3)
                 --model-dir <path>             Local model directory (skips download)
-                --encoder-precision <int8|int4> Encoder quantization (default: int8)
+                --encoder-precision <int8|int8-v2|int4> Encoder quantization (default: int8;
+                                               int8-v2 = int8-linear rebuild, avoids #760)
                 --language <code>              Language hint (e.g., en, de, fr, es)
                 --custom-vocab <file>          Apply vocabulary boosting in batch mode
                 --no-mel-context               Disable 80ms mel-context prepend for long-form batch ASR
diff --git a/Tests/FluidAudioTests/ASR/Parakeet/SlidingWindow/AsrModelsEncoderV2Tests.swift b/Tests/FluidAudioTests/ASR/Parakeet/SlidingWindow/AsrModelsEncoderV2Tests.swift
new file mode 100644
index 00000000..71f4e76b
--- /dev/null
+++ b/Tests/FluidAudioTests/ASR/Parakeet/SlidingWindow/AsrModelsEncoderV2Tests.swift
@@ -0,0 +1,59 @@
+import XCTest
+
+@testable import FluidAudio
+
+/// Opt-in Encoder_v2 (int8-linear, issue #760) precision selection.
+final class AsrModelsEncoderV2Tests: XCTestCase {
+
+    func testEncoderFileNames() {
+        XCTAssertEqual(ParakeetEncoderPrecision.int8.encoderFileName, "Encoder.mlmodelc")
+        XCTAssertEqual(ParakeetEncoderPrecision.int8V2.encoderFileName, "Encoder_v2.mlmodelc")
+        XCTAssertEqual(ParakeetEncoderPrecision.int4.encoderFileName, "EncoderInt4.mlmodelc")
+    }
+
+    func testVariantStringRoundTrip() {
+        // The download variant string must map back to the same precision so
+        // ModelHub's required-model set matches the file AsrModels loads.
+        for precision in ParakeetEncoderPrecision.allCases {
+            XCTAssertEqual(ParakeetEncoderPrecision(rawValue: precision.rawValue), precision)
+        }
+    }
+
+    func testRequiredModelsV3PerPrecision() {
+        // int8 stays the default and keeps requiring the original encoder;
+        // int8-v2 is strictly opt-in.
+        let int8 = ModelNames.ASR.requiredModelsV3(precision: .int8)
+        XCTAssertTrue(int8.contains("Encoder.mlmodelc"))
+        XCTAssertFalse(int8.contains("Encoder_v2.mlmodelc"))
+
+        let int8V2 = ModelNames.ASR.requiredModelsV3(precision: .int8V2)
+        XCTAssertTrue(int8V2.contains("Encoder_v2.mlmodelc"))
+        XCTAssertFalse(int8V2.contains("Encoder.mlmodelc"))
+    }
+
+    func testModelsExistChecksPrecisionSpecificEncoder() throws {
+        let parentDir = FileManager.default.temporaryDirectory
+            .appendingPathComponent("AsrModelsEncoderV2Tests-\(UUID().uuidString)")
+        let repoDir = parentDir.appendingPathComponent(Repo.parakeetV3.folderName)
+        defer { try? FileManager.default.removeItem(at: parentDir) }
+
+        for file in [
+            ModelNames.ASR.preprocessorFile,
+            ModelNames.ASR.decoderFile,
+            ModelNames.ASR.jointV3File,
+            ModelNames.ASR.encoderV2File,
+        ] {
+            try FileManager.default.createDirectory(
+                at: repoDir.appendingPathComponent(file), withIntermediateDirectories: true)
+        }
+        try Data("{}".utf8).write(
+            to: repoDir.appendingPathComponent(ModelNames.ASR.vocabularyFile))
+
+        // Only Encoder_v2 on disk: satisfies an explicit int8-v2 request but
+        // NOT the int8 default (which still requires Encoder.mlmodelc).
+        XCTAssertTrue(
+            AsrModels.modelsExist(at: repoDir, version: .v3, encoderPrecision: .int8V2))
+        XCTAssertFalse(
+            AsrModels.modelsExist(at: repoDir, version: .v3, encoderPrecision: .int8))
+    }
+}