blue-machines/tundra-mimi-encoder

Tundra Mimi Encoder

Speech encoder / audio codec weights used by Blue Machines Tundra speech-LLM training (encoder_type: mimi, discrete or continuous modes).

Source

• Base codec: kyutai/mimi
• Fine-tune focus: Indic languages (IndicVoices-R)
• File: mimi_final.pt — full model state_dict for the fine-tuned encoder / semantic codebook

Please cite / credit the original authors of Svara Mimi Indic v3 when using these weights.

Files

File	Description
mimi_final.pt	Fine-tuned Mimi weights (~184 MB)

Usage in this stack

model:
  encoder_type: mimi
  encoder_name_or_path: kyutai/mimi
  mimi_mode: discrete   # or continuous
  mimi_finetune_name_or_path: blue-machines/tundra-mimi-encoder
  mimi_num_quantizers: 8
  freeze_encoder: true

Typical load pattern (same as the upstream card):

import torch
from transformers import MimiModel  # or your local Mimi wrapper

# Load base architecture from kyutai/mimi, then:
state = torch.load("mimi_final.pt", map_location="cpu")
# model.load_state_dict(state, strict=False)