CSM-1B-MLX-fp16

Sesame CSM-1B (Conversational Speech Model) converted to MLX (fp16) for on-device speech generation on Apple Silicon. CSM is a Moshi-family codec model: a Llama-1B backbone predicts the zeroth Mimi codebook, a Llama-100M decoder predicts the rest. It clones a voice from a short reference clip.

Property	Value
Parameters	1.5B class (backbone 1B + decoder + heads)
Quantization	fp16
Size	~3.3 GB
Codec	Mimi, 32 codebooks @ 12.5 Hz, 24 kHz
Languages	English

Files

File	Description
model.safetensors	backbone + decoder + heads + embeddings (fp16)
config.json	architecture + per-module quantization map
mimi.safetensors	Kyutai Mimi codec (32 codebooks)
tokenizer.json	Llama BPE tokenizer

Performance (M-series, mlx-swift)

Metric	Value
RTF (int8)	~0.5–0.9 (faster than real time)
First-audio latency (streaming)	~120 ms

Quantization lowers memory bandwidth, so int8/int5 are both smaller and faster than fp16 on Apple GPUs.

Usage (Swift)

import CSM

let pipeline = try await CSMPipeline(directory: modelDir)
let audio = pipeline.synthesize(
    text: "Hello from on-device speech.",
    refAudio: referenceSamples,           // 24 kHz mono
    refText: "transcript of the reference",
    temperature: 0.9, topK: 50)

Known limitations

• English-focused; weak on other languages.
• End-of-audio occasionally clips the final word — pad the text with a short trailing phrase, or trim to the last voiced frame.
• Zero-shot cloning transfers voice identity but flattens prosody vs. a purpose-tuned reference.

Source

Converted from sesame/csm-1b (Apache-2.0).

Links

• speech-swift — Apple SDK
• Docs
• soniqo.audio
• Speech Studio — local speech generation & voice cloning
• blog