qwen3-8b-4bit @ main

Automated model bundle release for qwen3-8b-4bit.

Description: General-purpose local Qwen text model for chat, automation, and coding.

Use cases

• Local assistant chat
• Coding and automation
• Tool-enabled Flutter app experiments

Limitations

• Quality and speed depend on Apple Silicon memory pressure
• Tool calling requires a compatible runtime parser

Release metadata

• Runtime: mlx_lm
• Tasks: chat, code
• Source: mlx-community/Qwen3-8B-4bit@main
• Archive size: 4623841280 bytes
• Languages: multilingual
• Default parameters: {"max_tokens": 512, "temperature": 0.7, "top_p": 0.95}

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

qwen2-audio-7b-instruct-4bit @ main

Automated model bundle release for qwen2-audio-7b-instruct-4bit.

Description: Audio-first local Qwen model that accepts spoken input and returns text output.

Use cases

• Direct audio-question answering
• Voice assistant prototypes
• Audio understanding experiments

Limitations

• Audio chat support depends on model template and mlx runtime coverage
• May be slower than ASR plus text LLM pipeline

Release metadata

• Runtime: mlx_audio
• Tasks: chat, audio_input
• Source: mlx-community/Qwen2-Audio-7B-Instruct-4bit@main
• Archive size: 6574714880 bytes
• Languages: multilingual

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

qwen3-tts-12hz-1.7b-customvoice-4bit @ main

Automated model bundle release for qwen3-tts-12hz-1.7b-customvoice-4bit.

Description: Higher-quality Qwen3 TTS CustomVoice model with built-in multilingual speaker presets.

Use cases

• Stable preset voices for assistant playback
• Fast voice UX prototyping
• Multilingual local TTS experiments

Limitations

• Speaker presets work best in each speaker native language
• Optional instruction quality affects emotion and prosody

Release metadata

• Runtime: mlx_audio
• Tasks: text_to_speech, audio_output
• Source: mlx-community/Qwen3-TTS-12Hz-1.7B-CustomVoice-4bit@main
• Archive size: 2312130560 bytes
• Languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
• Output: {"media_type": "audio/wav", "extension": "wav", "supports_inline_playback": true}
• Default parameters: {"audio_format": "wav", "join_audio": true, "lang_code": "auto", "voice": "Ryan", "speed": 1.0}
• Voices: Vivian, Serena, Uncle Fu, Dylan, Eric, Ryan

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

qwen3-tts-12hz-0.6b-customvoice-4bit @ main

Automated model bundle release for qwen3-tts-12hz-0.6b-customvoice-4bit.

Description: Smaller Qwen3 TTS CustomVoice model with built-in multilingual speaker presets.

Use cases

• Stable preset voices for assistant playback
• Fast voice UX prototyping
• Multilingual local TTS experiments

Limitations

• Speaker presets work best in each speaker native language
• Optional instruction quality affects emotion and prosody

Release metadata

• Runtime: mlx_audio
• Tasks: text_to_speech, audio_output
• Source: mlx-community/Qwen3-TTS-12Hz-0.6B-CustomVoice-4bit@main
• Archive size: 1693675520 bytes
• Languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian
• Output: {"media_type": "audio/wav", "extension": "wav", "supports_inline_playback": true}
• Default parameters: {"audio_format": "wav", "join_audio": true, "lang_code": "auto", "voice": "Ryan", "speed": 1.0}
• Voices: Vivian, Serena, Uncle Fu, Dylan, Eric, Ryan

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

z-image-turbo-mflux-4bit @ main

Automated model bundle release for z-image-turbo-mflux-4bit.

Description: Fast Z-Image Turbo mflux 4bit model for low-latency local text-to-image generation.

Use cases

• Fast local text-to-image generation
• Photorealistic prompt previews
• Offline creative workflows on Apple Silicon

Limitations

• License metadata should be reviewed before redistribution/use in products
• Very low step counts can trade detail for latency

Release metadata

• Runtime: mflux
• Tasks: image_generation
• Source: filipstrand/Z-Image-Turbo-mflux-4bit@main
• Archive size: 5911244800 bytes
• Languages: English prompts
• Output: {"media_type": "image/png", "extension": "png", "default_size": "1024x1024"}
• Default parameters: {"width": 1024, "height": 1024, "steps": 9, "guidance": 3.5}

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

whisper-tiny-asr-4bit @ main

Automated model bundle release for whisper-tiny-asr-4bit.

Description: Tiny multilingual Whisper MLX-Audio speech-to-text model for fast Russian and English dictation tests.

Use cases

• Fast voice input for voice-to-voice pipeline tests
• Russian and English ASR smoke tests
• Offline dictation prototypes

Limitations

• Lowest accuracy Whisper tier
• No audio generation; pair with a TTS model for voice output

Release metadata

• Runtime: mlx_audio
• Tasks: speech_to_text, audio_input
• Source: mlx-community/whisper-tiny-asr-4bit@main
• Archive size: 26490880 bytes
• Languages: English, Russian, multilingual
• Output: {"media_type": "text/plain", "format": "text"}
• Default parameters: {"task": "transcribe", "language": "auto", "temperature": 0.0, "word_timestamps": false}

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

whisper-small-asr-4bit @ main

Automated model bundle release for whisper-small-asr-4bit.

Description: Compact multilingual Whisper MLX-Audio model for better Russian and English transcription under 150 MB.

Use cases

• Better-quality voice input for local assistants
• Russian and English ASR validation
• Offline transcription and translation experiments

Limitations

• Slower than tiny/base variants
• No audio generation

Release metadata

• Runtime: mlx_audio
• Tasks: speech_to_text, audio_input
• Source: mlx-community/whisper-small-asr-4bit@main
• Archive size: 143616000 bytes
• Languages: English, Russian, multilingual
• Output: {"media_type": "text/plain", "format": "text"}
• Default parameters: {"task": "transcribe", "language": "auto", "temperature": 0.0, "word_timestamps": false}

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

whisper-base-asr-4bit @ main

Automated model bundle release for whisper-base-asr-4bit.

Description: Small multilingual Whisper MLX-Audio model balancing speed and Russian/English transcription quality.

Use cases

• Russian and English speech-to-text
• Voice assistant ASR stage
• Low-memory offline transcription

Limitations

• Lower accuracy than Whisper Small/Large variants
• No audio generation

Release metadata

• Runtime: mlx_audio
• Tasks: speech_to_text, audio_input
• Source: mlx-community/whisper-base-asr-4bit@main
• Archive size: 46694400 bytes
• Languages: English, Russian, multilingual
• Output: {"media_type": "text/plain", "format": "text"}
• Default parameters: {"task": "transcribe", "language": "auto", "temperature": 0.0, "word_timestamps": false}

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

voxtral-4b-tts-2603-4bit @ main

Automated model bundle release for voxtral-4b-tts-2603-4bit.

Description: Multilingual Voxtral TTS model with 20 built-in voice presets and faster-than-real-time MLX output.

Use cases

• Stable preset TTS voices for voice-to-voice pipeline output
• Multilingual assistant speech playback
• Comparing Qwen3-TTS and Voxtral voice quality

Limitations

• Non-commercial source license
• Text-to-speech only; pair with ASR or STS for full voice workflows

Release metadata

• Runtime: mlx_audio
• Tasks: text_to_speech, audio_output
• Source: mlx-community/Voxtral-4B-TTS-2603-mlx-4bit@main
• Archive size: 2542929920 bytes
• Languages: English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic, Hindi
• Output: {"media_type": "audio/wav", "extension": "wav", "sample_rate_hz": 24000, "supports_inline_playback": true}
• Default parameters: {"voice": "casual_male", "audio_format": "wav", "sample_rate": 24000}
• Voices: Casual Male, Casual Female, Cheerful Female, Neutral Male, Neutral Female, French Male

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.

voxcpm2-4bit @ main

Automated model bundle release for voxcpm2-4bit.

Description: Multilingual MLX text-to-speech and voice cloning model for local assistant voices.

Use cases

• Spoken assistant responses
• Voice cloning with short reference audio and transcript
• Local TTS playback in Flutter chat
• Voice UX prototyping

Limitations

• Voice controls vary by model family
• Generated speech quality depends on prompt and runtime support
• Use a clean 3-10 second reference clip and transcript for stable voice cloning
• Requires a GitHub mlx-audio build because PyPI mlx-audio 0.4.3 does not include VoxCPM2 yet

Release metadata

• Runtime: mlx_audio
• Tasks: text_to_speech, audio_output
• Source: mlx-community/VoxCPM2-4bit@main
• Archive size: 2302095360 bytes
• Languages: multilingual
• Output: {"media_type": "audio/wav", "extension": "wav", "supports_inline_playback": true}
• Default parameters: {"audio_format": "wav", "join_audio": true, "lang_code": "en", "cfg_scale": 2.0, "ddpm_steps": 7, "max_tokens": 2000}
• Voices: Default VoxCPM2

This release includes model_metadata.json, release_metadata.json, manifest.source.yaml, and chunked model archive assets.