diff --git a/README.md b/README.md
index 13488d4..45943b2 100644
--- a/README.md
+++ b/README.md
@@ -51,12 +51,16 @@ written is still readable.
 
 ## Transcription
 
-Built in, on-device, automatic. The default engine is **Parakeet TDT 0.6B v2**
-(English) via [FluidAudio](https://github.com/FluidInference/FluidAudio)'s
-Core ML port — roughly 20 seconds per hour of audio on Apple Silicon. Models
-(~600 MB) download once on first transcription; `quill doctor` tells you
-whether they're already cached so you're never downloading after an important
-meeting.
+Built in, on-device, automatic. The default engine is **Parakeet TDT 0.6B v3**
+via [FluidAudio](https://github.com/FluidInference/FluidAudio)'s Core ML port —
+roughly 20 seconds per hour of audio on Apple Silicon. v3 is the multilingual
+model (25 European languages plus Japanese) and detects the spoken language on
+its own, so there's nothing to configure for a non-English meeting.
+
+Set `transcription.model` to `"v2"` for the English-only model, which has
+marginally higher recall on English. Models (~600 MB) download once on first
+transcription; `quill doctor` tells you whether they're already cached so
+you're never downloading after an important meeting.
 
 Each track is transcribed separately, shifted by its start offset so both
 share one clock, and merged by timestamp. Jobs run in a serial queue — you can
@@ -75,7 +79,7 @@ Optional, at `~/.config/quill/config.json`:
 ```json
 {
   "recordings_dir": "~/Recordings",
-  "transcription": { "enabled": true, "engine": "parakeet" },
+  "transcription": { "enabled": true, "engine": "parakeet", "model": "v3" },
   "on_stop": "my-hook"
 }
 ```
@@ -83,6 +87,9 @@ Optional, at `~/.config/quill/config.json`:
 - `recordings_dir` — where sessions land. Resolution order: `--out` flag >
   config > `~/Recordings`.
 - `transcription.enabled` — set `false` to just record.
+- `transcription.model` — parakeet model version: `"v3"` (default,
+  multilingual, self-detecting) or `"v2"` (English-only, marginally higher
+  recall on English). An unrecognized value warns and falls back to v3.
 - `mic_voice_processing` — Apple's echo cancellation on the mic (default off).
   Set `true` when recording meetings through the speakers, so playback doesn't
   bleed into the mic track and get transcribed twice as "me". The trade: while
@@ -121,7 +128,8 @@ quill install --uninstall
   per-process picker if it bothers you).
 - If recordings come out silent, check System Settings → Privacy & Security →
   Screen & System Audio Recording.
-- Parakeet v2 is English-only. Other languages will come with the Whisper
-  engine.
+- Parakeet v2 (`"model": "v2"`) is English-only — it will happily return
+  English-looking nonsense for other languages rather than failing. v3 is the
+  default for that reason.
 - The binary embeds its Info.plist (`__TEXT,__info_plist`) so TCC can
   attribute permissions to quill itself when running as a LaunchAgent.
diff --git a/Sources/quill/Config.swift b/Sources/quill/Config.swift
index 5db668f..df0f4c8 100644
--- a/Sources/quill/Config.swift
+++ b/Sources/quill/Config.swift
@@ -44,6 +44,13 @@ enum Config {
         transcription()?["engine"] as? String ?? "parakeet"
     }
 
+    /// Configured model for the engine. For parakeet: "v3" (multilingual,
+    /// default) or "v2" (English-only, marginally higher recall on English).
+    /// The engine warns and falls back for anything it doesn't recognize.
+    static func transcriptionModel() -> String {
+        transcription()?["model"] as? String ?? "v3"
+    }
+
     private static func transcription() -> [String: Any]? {
         load()?["transcription"] as? [String: Any]
     }
diff --git a/Sources/quill/Doctor.swift b/Sources/quill/Doctor.swift
index 8af8125..ec1a3dd 100644
--- a/Sources/quill/Doctor.swift
+++ b/Sources/quill/Doctor.swift
@@ -86,8 +86,9 @@ enum DoctorReport {
                 remediation: nil
             )
         }
-        let cache = AsrModels.defaultCacheDirectory(for: .v2)
-        if AsrModels.modelsExist(at: cache, version: .v2) {
+        let version = ParakeetEngine.configuredVersion()
+        let cache = AsrModels.defaultCacheDirectory(for: version)
+        if AsrModels.modelsExist(at: cache, version: version) {
             return Check(name: "transcription", status: .ok, remediation: nil)
         }
         return Check(
diff --git a/Sources/quill/Transcription/ParakeetEngine.swift b/Sources/quill/Transcription/ParakeetEngine.swift
index ff6cad5..f51c360 100644
--- a/Sources/quill/Transcription/ParakeetEngine.swift
+++ b/Sources/quill/Transcription/ParakeetEngine.swift
@@ -2,10 +2,14 @@ import AVFoundation
 import FluidAudio
 import Foundation
 
-/// Parakeet TDT 0.6B v2 (English) via FluidAudio's Core ML port. Models
-/// download once into FluidAudio's managed cache (~600 MB); after that,
-/// transcription runs entirely on-device at roughly 20 seconds per hour of
-/// audio on Apple Silicon.
+/// Parakeet TDT 0.6B via FluidAudio's Core ML port. Models download once into
+/// FluidAudio's managed cache (~600 MB); after that, transcription runs
+/// entirely on-device at roughly 20 seconds per hour of audio on Apple
+/// Silicon.
+///
+/// Two model versions, selected with `transcription.model`: v3 is multilingual
+/// (25 European languages plus Japanese) and detects the spoken language
+/// itself; v2 is English-only with marginally higher recall on English.
 actor ParakeetEngine: TranscriptionEngine {
     enum EngineError: Error, CustomStringConvertible {
         case notPrepared
@@ -21,14 +25,37 @@ actor ParakeetEngine: TranscriptionEngine {
         }
     }
 
+    /// The configured model version, warning and falling back rather than
+    /// silently transcribing with one the user didn't ask for. Shared with
+    /// `quill doctor` so the cache check can't drift from what we download.
+    static func configuredVersion() -> AsrModelVersion {
+        switch Config.transcriptionModel() {
+        case "v3": return .v3
+        case "v2": return .v2
+        case let other:
+            FileHandle.standardError.write(Data(
+                "warning: unknown parakeet model \"\(other)\" — using v3\n".utf8
+            ))
+            return .v3
+        }
+    }
+
     nonisolated let name = "parakeet"
-    nonisolated let model = "parakeet-tdt-0.6b-v2-coreml"
+    nonisolated let model: String
+    private let version: AsrModelVersion
 
     private var manager: AsrManager?
 
+    init(version: AsrModelVersion = ParakeetEngine.configuredVersion()) {
+        self.version = version
+        self.model = version == .v2
+            ? "parakeet-tdt-0.6b-v2-coreml"
+            : "parakeet-tdt-0.6b-v3-coreml"
+    }
+
     func prepare() async throws {
         guard manager == nil else { return }
-        let models = try await AsrModels.downloadAndLoad(version: .v2)
+        let models = try await AsrModels.downloadAndLoad(version: version)
         let manager = AsrManager()
         try await manager.loadModels(models)
         self.manager = manager
@@ -69,7 +96,7 @@ actor ParakeetEngine: TranscriptionEngine {
     }
 
     /// Group word timings into readable segments: break on sentence-ending
-    /// punctuation (parakeet v2 emits punctuation), a silence gap, or a hard
+    /// punctuation (both parakeet versions emit it), a silence gap, or a hard
     /// length cap so a run-on speaker still wraps.
     private static func segments(from words: [WordTiming]) -> [TranscriptSegment] {
         var out: [TranscriptSegment] = []