File size: 778 Bytes
4cef0dd | 1 2 3 4 5 6 7 8 9 10 11 12 | # T8 modifications
Powered by MiniMax H3. Derived from Viggle/Meridian at revision `2083d059d8544ff7eaaf86966b83e4964a904737`.
1. Paired Meridian DMD LoRA merged into the Meridian-specific finetuned transformer at strength1.0; not a stock H3 checkpoint.
2. Converted keys/layout for ComfyUI native loading.
3. Converted250 main-block linear layers to native ConvRot INT8, retaining nonquantized tensors;535 base tensors plus500 quantization auxiliaries,1035 keys.
4. No new training. Do not apply the paired DMD again.
The separate original Omega1B512 PT file is unchanged, unquantized, and has a separate FAIR license.
Original teacher shards, H3 VAE, account credentials and review media are not included. See conversion_manifest.json and the upstream modification notice.
|