Ultravox v0.5 Llama 3.2 1B (Core ML Native)

Private Core ML release of Ultravox v0.5 (fixie-ai/ultravox-v0_5-llama-3_2-1b) for TranslateBlue on iOS and macOS.

Architecture

  • Audio Encoder: Whisper-large-v3-turbo encoder (128 mel bins)
  • Projector: SwiGLU 2-layer MLP (1280 → 2048)
  • LLM Decoder: Llama-3.2-1B-Instruct with stateful KV cache (UltravoxLlamaKVCache.mlpackage)

Artifacts

Path Role Notes
AudioEncoder.mlpackage Encoder source On-device compile fallback (macOS only)
AudioEncoder.mlmodelc Encoder (macOS) Do not load on iOS
AudioEncoder-ios.mlmodelc Encoder (iOS) Pre-compiled; preferred on iPhone/iPad
UltravoxLlamaKVCache.mlpackage Decoder Stateful Core ML; 2.47 GB weight.bin
tokenizer.json BPE tokenizer Legacy string merges (280,147 entries); required for inference
tokenizer_config.json Tokenizer metadata Not sufficient alone — tokenizer.json is required
ultravox_llama_kvcache_manifest.json Decoder manifest Input/output/state names

Platform-specific compiled encoder bundles are also published in sibling repos:

Requirements

  • iOS 18.0+ or macOS 15.0+
  • Physical device required for end-to-end validation — the decoder is a stateful Core ML model (computeUnits = .cpuAndGPU; ANE unsupported today).

iOS Simulator is not supported

UltravoxLlamaKVCache.mlpackage fails to compile on the iOS Simulator. Observed error:

Espresso exception: "Invalid state": MpsGraph backend validation on incompatible OS
compilation failed — Failed to build the model execution plan … model.mil … error code: -14

Use a physical iPhone or iPad for decoder load and speech-to-text translation tests.

Tokenizer format

tokenizer.json uses the legacy BPE merges string format (["Ġ Ġ", …]), not the tokenizers 0.20+ pair format ([["Ġ","Ġ"], …]). Some Rust tokenizers loaders silently return null on the new format. Byte-level BPE encodes spaces as Ġ, so the string join is unambiguous and reversible.

Tasks

  • Direct Speech-to-Text Translation (S2TT)
  • Context-Aware Speech Recognition (ASR)
  • Voice Conversation / QA
Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support