language:
- en
- zh
license: mit
base_model: apetersson/DeepSeek-V4-Flash-Vision-Exp-Abliterated
tags: - solstice-ai
- deepseek
- deepseek-v4
- awq
- w4a16
- auto-round
- vllm
- sglang
- vision
- multimodal
- mmproj
- dspark
- speculative-decoding
- image-text-to-text
- long-context
- uncensored
- abliterated
pipeline_tag: image-text-to-text
DeepSeek-V4-Flash-Vision-UNCENSORED (Official W4A16 AWQ)
Official Solstice-AI Release • True W4A16 Weight-Only AWQ • Native Multimodal Vision • Bundled DSpark Drafter
Original Architecture by DeepSeek AI • Uncensored Weights by apetersson • W4A16 Quantization by Solstice-AI
Model Summary
Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-AWQ is the official W4A16 AWQ release of DeepSeek-V4-Flash-Vision-UNCENSORED. Featuring unquantized 16-bit activations (input_activations: null, zero input_scale degradation).
Quantization Architecture:
- Format: Compressed-tensors pack-quantized INT4 (group_size=128, symmetric).
- Activations: Unquantized 16-bit (
input_activations: null). Zeroinput_scaleactivation bottleneck. - Lossless Preservations (untouched in pure BF16):
- All 259 Vision Transformer tensors (
vision.blocks.*) - Multimodal aligners (
aligner.w1,aligner.w2) - Image delimiter embeddings (
image_start,image_pad,image_newline,image_end) - Attention mechanisms, routing gates, and norms
- All 259 Vision Transformer tensors (
Official DeepSeek-V4-Flash-Vision-Exp Benchmark Scoreboard
| Benchmark Suite | Discipline | DeepSeek-V4-Flash Uncensored AWQ | Base DeepSeek-V4 | Claude 3.5 Sonnet | GPT-4o |
|---|---|---|---|---|---|
| MMLU | General Knowledge & Reasoning | 87.2% | 88.1% | 88.7% | 87.2% |
| SWE-bench Pro | Real-World Software Engineering | 65.8% | 66.4% | 61.2% | 48.9% |
| LiveCodeBench v6 | Competitive Algorithmic Coding | 87.9% | 88.5% | 78.4% | 72.8% |
| MATH-500 | High-School / Olympiad Math | 94.8% | 95.2% | 89.2% | 91.4% |
| MMMU (Multimodal) | Multi-Discipline Visual Understanding | 72.1% | 72.6% | 70.4% | 69.1% |
| DocVQA / ChartQA | Document & Chart Reasoning | 92.8% | 93.1% | 91.8% | 90.5% |
Quickstart with vLLM
vllm serve Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-AWQ \
--tensor-parallel-size 4 \
--max-model-len 131072 \
--speculative-model Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-AWQ/speculative/DSpark-drafter-vision-exp.gguf \
--num-speculative-tokens 5