polyvox-models

Mirror of the model files Polyvox downloads at runtime. Every file here is an INT8 ONNX export produced in-house from an upstream checkpoint that permits redistribution. The paths match the relative paths the app stores the files under, so one base URL swap covers every key.

This repo exists so mainland China users can fetch these weights. Their previous only home was a GitHub release asset, which the Great Firewall makes unreliable, and no HuggingFace proxy could help because there was no HuggingFace original to proxy.

Contents

Path	Size	Upstream	License
m2m100/encoder_model_int8.onnx	612 MB	facebook/m2m100_1.2B	MIT
m2m100/decoder_model_int8.onnx	835 MB	same	MIT
m2m100/tokenizer.json	7.7 MB	same	MIT
fireredlid/fireredlid_encoder/fireredlid_encoder.onnx	1.2 GB	FireRedTeam FireRedLID	Apache-2.0
fireredlid/fireredlid_decoder/fireredlid_decoder.onnx	177 MB	same	Apache-2.0
fireredlid/fireredlid.json	8 KB	same (vocab + CMVN sidecar)	Apache-2.0
fireredasr/fireredasr_encoder/fireredasr_encoder.onnx	1.2 GB	FireRedTeam FireRedASR2-AED	Apache-2.0
fireredasr/fireredasr_decoder/fireredasr_decoder.onnx	481 MB	same	Apache-2.0
fireredasr/fireredasr.json	92 KB	same (vocab + CMVN sidecar)	Apache-2.0
fireredvad_aed.onnx	2.3 MB	FireRedTeam FireRedVAD	Apache-2.0

The two sidecar JSON files carry the vocabulary and the CMVN statistics. The graphs are unusable without them, because upstream ships those as a Kaldi cmvn.ark plus a dict.txt that no ONNX runtime reads.

Changes made to the upstream weights

Apache-2.0 section 4(b) requires stating this, and it is useful anyway:

• FireRed (LID, ASR2-AED, VAD): upstream publishes no ONNX for any module. Each repo ships model.pth.tar, a Kaldi cmvn.ark and a dict.txt. These graphs were exported here, split at the encoder/decoder boundary with fixed input shapes, then INT8 dynamic quantization applied to the MatMuls. The VAD is an opset-17 re-export with CMVN compiled into the graph.
• M2M-100 1.2B: exported to ONNX, split at the encoder/decoder boundary, INT8 dynamic quantization on the MatMuls.

Attribution

• FireRedLID, FireRedASR2-AED and FireRedVAD are by FireRedTeam, Apache-2.0 (arXiv 2603.10420). Apache-2.0 permits commercial use, modification and redistribution of derivative works with notice retention, which is what this repo does.
• M2M-100 is by Meta AI, MIT. The MIT notice and Meta copyright travel with the Polyvox build on its third-party notices screen.

Not here, deliberately

Two families Polyvox uses are absent because their licenses forbid us redistributing them, and no mirror request will change that:

• NLLB-200 (Xenova/nllb-200-distilled-600M) is CC-BY-NC-4.0. Polyvox never bundles, hosts, mirrors or serves a copy. Users fetch it from the upstream repo on an explicit button press.
• Sortformer INT8 is NVIDIA-derived and the attribution surface for the INT8 conversion's lineage is still open.