Drama 3 for ComfyUI β€” fp8 experts, bf16 dense

A conversion of fishaudio/drama-3 for the ComfyUI-Drama3 node pack.

  • Copied from: https://huggingface.co/fishaudio/drama-3 β€” revision 78e5ab79e49a, the weights uploaded 2026-10-29 (15 shards; the earlier 8-shard files were removed upstream that day).
  • Script used: tools/convert.py from https://github.com/PedroMarinhoDev/ComfyUI-Drama3 python tools/convert.py model <drama-3 folder> drama3_fp8experts_bf16.safetensors --format fp8
  • What is in the file: the routed experts are the checkpoint's own block-FP8 tensors (128x128 scales) copied byte for byte; every other linear keeps the checkpoint's bf16. Gate/up projections are fused; the LM head is trimmed to the rows generation samples from (1 end-of-speech + 4096 semantic tokens).
  • How to use: put the file in ComfyUI/models/diffusion_models/ and load it with the Drama3Loader node. Pair it with drama3_codec.safetensors in ComfyUI/models/vae/ and Qwen3-ASR in ComfyUI/models/asr/ (both from the pack's own Hub repos).
  • File: drama3_fp8experts_bf16.safetensors β€” 38,578,417,928 bytes, 912 tensors, sha256 cfd27b526b35a66729ffd1d6966bf13f7856f4c59f917497eaded6ecb0b888de.

Note: the prebuilt file in PedroMarinhoDev/Drama-3-ComfyUI comes from an earlier revision of the base model. This file matches the 2026-10-29 weights, and its tensors were verified byte-identical to that checkpoint.

License: Fish Audio's (see the fishaudio/drama-3 model card and the license file in this repo).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for zzb1420/Drama-3-ComfyUI

Finetuned
(8)
this model