moshiko-7b-bf16 (Vokra GGUF)

Speech to speech model, converted to the Vokra GGUF format for Vokra, a zero-dependency speech-AI inference runtime.

This is a conversion, not a new model. The weights are the upstream ones; Vokra re-packages them so its runtime can memory-map them directly. Credit for the model belongs upstream — see Source below.

Files

File	Size	SHA-256
moshi-7b-bf16-full.gguf	14663.7 MB	082b9c24c08adc5a5cb10099be261fbaa2599db261d55efc872162f349951d3e

Usage

# Download (any HTTP client works — the file is a plain GGUF)
curl -L -o moshi-7b-bf16-full.gguf \
  https://huggingface.co/vokra/moshiko-7b-bf16/resolve/main/moshi-7b-bf16-full.gguf

# This artifact is consumed by another engine rather than run
# standalone — see the Vokra docs for the S2S pipeline.

Provenance

Field	Value
Architecture	moshi
Tensors	355
Upstream source	https://huggingface.co/kyutai/moshiko-pytorch-bf16
Upstream licence	CC-BY-4.0
Licence class	attribution-required
Registry model id	moshi
Vokra GGUF schema	(pre-stamping)
Converted by	(unrecorded)

Every row above is read out of this file's own vokra.* metadata, so the card cannot claim something the artifact does not carry.

Licence

The weights are distributed under CC-BY-4.0, unchanged from upstream. Conversion does not alter the licence, and your obligations run to the upstream author.

Attribution required

This application uses the Moshi model (Helium temporal transformer + depformer + Mimi codec) by Kyutai. Moshi weights are licensed under CC-BY 4.0 (attribution required; commercial use permitted). Copyright (c) Kyutai / Moshi authors. Source: https://github.com/kyutai-labs/moshi / https://huggingface.co/kyutai/moshiko-pytorch-bf16

This licence obliges you to display the attribution above when you ship something built on these weights. Vokra surfaces it at runtime via vokra_model_attribution (C ABI) and a CLI banner.

⚠️ Not real-time on Vokra's current runtime

Moshi's premise is real-time full-duplex dialogue (~200 ms on GPU). On Vokra's current CPU runtime this converted weight runs at RTF ≈ 1000 — it is a faithful BF16 parity/functional artifact, not a usable real-time model yet. GPU kernels for the Helium temporal + per-step-weight depformer + Mimi chain are a follow-up. Full-duplex use also needs an acoustic echo canceller (AEC) or the model self-echoes.

Verifying this file

shasum -a 256 moshi-7b-bf16-full.gguf
# expect: 082b9c24c08adc5a5cb10099be261fbaa2599db261d55efc872162f349951d3e