Prism Q6 (blocks-q6-v1)
A first-attempt, zero-optimization 6-bit quantization of Tencent Hunyuan's Prism image-to-video-and-audio model. Produced and tested on a single RTX 5090. Shared for experimentation and hardware comparisons — quality is experimental (see Known limitations).
| Format | prism-custom-q6 — block Linear weights packed at 6 bits, group size 64, FP32 scales; all other weights retained |
| Size | 27.1 GB across 52 Safetensors shards (source checkpoint: 65.3 GB BF16) |
| Source weights | FrancisRing/Prism @ a4ea8f5f6a5d3b71df04a048ed05c4e916f8595e |
| Source code | Tencent-Hunyuan/Prism @ eb943ff6079a8ea261247e965f9556703109c1c0 |
| Manifest SHA-256 | 74eb0f622587458b4eb7c4f18d9de4ea95ee56af85141cd3824cc778602963e1 |
Announcement: https://x.com/atomtanstudio/status/2107415066299809977
Measured on RTX 5090 (single GPU, CPU offload)
- 480p, 24 fps, ~5 s clip (121 frames), 50 steps: ~30 minutes, runs in 16 GB VRAM (peak CUDA allocation 11.7 GB / reservation 13.3 GB), peak process RSS ~44 GB host RAM.
- Native 720p, 61 frames, 50 steps: ~31-34 minutes.
Repository layout
manifest.json— versioned Q6 manifest: shard SHA-256s, tensor map, per-module quantization records, provenance.model-000NN.safetensors— the 52 quantized shards.receipts/— per-shard conversion receipts produced during quantization.loader/prism_quant/— the loader/conversion package (quant_loader.pyverifies every shard hash before loading;quant.pyimplements the Q6 format andQ6Linear;runtime.pyis the research runtime integration it was validated with).loader/convert.py— resumable script that regenerates these shards from the official BF16 checkpoint.
Using it
The format is custom Q6, not GGUF — stock Prism checkpoints loaders will
not read it directly. Use prism_quant with the Prism inference code:
from prism_quant.quant_loader import read_manifest, verify_shards
path, manifest = read_manifest("Prism-Q6") # validates format + shard hashes
verify_shards(path, manifest) # full SHA-256 verification
Requires PyTorch and safetensors. Shards are loaded block-wise with CPU
offload; in testing the launcher required at least 45 GiB free host RAM and
24 GiB free GPU memory before starting a run.
Known limitations
This is the alpha blocks-q6-v1 artifact: a first attempt with zero
optimizations. In completed test clips: restricted articulation, mouth and
teeth artifacts, lip sync unverified, and audio that can sound distorted.
Exact ASR transcription of prompts was verified, but that is not a quality
approval. Treat this as an experimental preview, not a finished preset.
License
MIT, inherited from the base Prism release. This repository contains a quantized derivative of the FrancisRing/Prism weights and adds no additional restrictions.