Downloads · 30 days
14
47% of all-time downloads
Dark-Scarlett-v1.0-31B-W4A16-multilingual
Ful3/Dark-Scarlett-v1.0-31B-W4A16-multilingual
Dark-Scarlett-v1.0-31B-W4A16-multilingual is a machine learning model from Ful3. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for vllm.
4-bit weight-only quantization of ReadyArt/Dark-Scarlett-v1.0-31B (Gemma 4 31B) made with llm-compressor, for vLLM.
All-time downloads
30
Public
Parameters
6.6B
20.5 GB on disk
Likes
0
Public
Repo makeup
Click a slice to open those files.
.safetensors20.4 GB · 100%
Parameter types
How the weights are stored.
I3229.3B · 94%
Base models
Model card
From the Hugging Face model README
Dark-Scarlett-v1.0-31B — W4A16 GPTQ, multilingual calibration
4-bit weight-only quantization of ReadyArt/Dark-Scarlett-v1.0-31B (Gemma 4 31B) made with llm-compressor, for vLLM.
- Attention and MLP projections: 4-bit GPTQ, weight-only (W4A16), group size 32
- Embeddings, lm_head, vision tower: BF16 (unquantized)
- Size on disk: ~20 GiB; fits two 32k sequences on a 32 GB card
Multilingual calibration. GPTQ was calibrated on 320 conversations at 4096 tokens: 96 English roleplay conversations plus 224 self-play roleplay conversations written by the BF16 model itself in Spanish, German, Italian, French, Korean, Japanese (28 each), Czech (20), Polish, Portuguese and Russian (12 each). An English-only calibration of the same recipe leaked English words into other languages and dropped Spanish accents; this build does not. On a 92-reply companion-chat check in 8 languages it matches the BF16 model (no English leaks, no degenerate replies).
vllm serve <this-repo> --language-model-only --kv-cache-dtype fp8
License and usage terms follow the base model (Gemma terms).
- Library
- vllm
- Type
- gemma4
- Created
- Sep 20, 2026
- Updated
- Sep 20, 2026