feat: Experimental Voxtral audio support

Add experimental support for Voxtral audio models:
- Add "voxtral" to AUDIO_MODEL_TYPES
- Increase audio limit: 5MB → 50MB (CLI + Server)
  - Gemma-3n: ~30s duration (188 tokens @ 6.25 tokens/s)
  - Voxtral: >10min duration (>2000 tokens)
  - Token count is real constraint, file size is sanity check

Requirements:
- mlx-vlm with Voxtral support (upstream merge pending)
- mistral-common for tekken tokenizer

Status: EXPERIMENTAL - pending mlx-vlm upstream integration

mlxk2/cli.py

@@ -498,9 +498,11 @@ def main():
                         print_result(result, None, True if args.json else False)
                         sys.exit(1)
                     data = aud_path.read_bytes()
-                    # 5MB limit for audio (~2-3 min at 16kHz mono; token count is the real constraint)
-                    if len(data) > 5 * 1024 * 1024:
-                        result = handle_error("CommandError", f"Audio file too large (>5MB): {audio_path}")
+                    # 50MB limit for audio (~15 min at 16kHz mono)
+                    # Note: Gemma-3n ~30s (token limit), Voxtral >10min (larger token capacity)
+                    # Token count is the real constraint, file size is just a sanity check
+                    if len(data) > 50 * 1024 * 1024:
+                        result = handle_error("CommandError", f"Audio file too large (>50MB): {audio_path}")
                         print_result(result, None, True if args.json else False)
                         sys.exit(1)
                     audio_inputs.append((aud_path.name, data))

mlxk2/core/capabilities.py

@@ -80,6 +80,7 @@ VISION_MODEL_TYPES = frozenset({
 AUDIO_MODEL_TYPES = frozenset({
     "gemma3n",        # Google Gemma 3n (Vision + Audio + Text)
     "gemma3n_audio",  # Audio encoder subcomponent
+    "voxtral",        # Voxtral mini (Audio + Text) - EXPERIMENTAL (pre-mlx-vlm merge)
 })
 
 

mlxk2/tools/vision_adapter.py

@@ -23,8 +23,10 @@ MAX_SAFE_CHUNK_SIZE = 5  # Empirically tested stable (5 images @ ~50MB total)
 SUPPORTED_MIME_TYPES = frozenset({"jpeg", "jpg", "png", "gif", "webp"})
 
 # Audio limits (ADR-019 Phase 4)
-# 5MB limit matches CLI (~2-3 min at 16kHz mono, token count is the real constraint)
-MAX_AUDIO_SIZE_BYTES = 5 * 1024 * 1024  # 5 MB per audio file
+# 50MB limit for audio (~15 min at 16kHz mono)
+# Note: Gemma-3n ~30s (token limit), Voxtral >10min (larger token capacity)
+# Token count is the real constraint, file size is just a sanity check
+MAX_AUDIO_SIZE_BYTES = 50 * 1024 * 1024  # 50 MB per audio file
 # mlx-vlm natively supports WAV and MP3 (verified in mlx-vlm README)
 SUPPORTED_AUDIO_FORMATS = frozenset({"wav", "mp3", "mpeg"})