cstr/kyutai-stt-2.6b-en-GGUF

Kyutai STT 2.6B English — GGUF

GGUF conversion of kyutai/stt-2.6b-en for the CrispASR kyutai-stt backend.

High-quality English ASR from the Kyutai/Moshi team. Causal transformer LM conditioned on Mimi audio codec features — streaming-capable architecture.

Architecture

• Mimi encoder: SEANet CNN + transformer + RVQ (audio -> discrete tokens)
• Causal LM: 16-layer transformer (2048-dim, RoPE, SwiGLU, RMSNorm)
• Input: 16 kHz mono audio
• Output: English text transcription

Files

File	Size	Description
kyutai-stt-2.6b-q4_k.gguf	1.5 GB	Q4_K quantized (recommended)
kyutai-stt-2.6b-q8_0.gguf	2.8 GB	Q8_0 quantized
kyutai-stt-2.6b-f16.gguf	5.1 GB	F16 full precision
kyutai-stt-2.6b-ref.gguf	4.0 MB	Diff harness reference

Usage

# Auto-download:
crispasr --backend kyutai-stt -m auto --auto-download -f audio.wav

# Local model:
crispasr --backend kyutai-stt -m kyutai-stt-2.6b-q4_k.gguf -f audio.wav -osrt

License

Apache 2.0 (same as the original model).

Credits

• Original model: Kyutai
• GGUF conversion & runtime: CrispASR