Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-AWQ-DSpark

DeepSeek-V4-Flash-Vision-UNCENSORED (W4A16 AWQ INT4 GEMM (DSpark))

Official Solstice-AI Release • Lossless Pure BF16 Multimodal Vision Tower • Native 1M YaRN Context • Bundled Native DSpark Speculative Drafter

Original Architecture by DeepSeek AI • Uncensored Weights by orcarouter • Quantization & DSpark Optimization by Solstice-AI

---

Model Summary

Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-AWQ-DSpark is the official W4A16 AWQ INT4 GEMM (DSpark) release of DeepSeek-V4-Flash-Vision-UNCENSORED (305B total parameters, 256 routed MoE experts, ~18B active per token).

Quantization Architecture:

• Format: Compressed-tensors W4A16 AWQ (group_size=128, symmetric).
• Activations: Unquantized 16-bit (W4A16 GEMM).
• Lossless Preservations: All 263 Vision Transformer & aligner tensors in untouched BF16.
• Router Preservation: MoE gating projections kept unquantized.
• Speculative Drafter: Bundled pre-aligned DSpark drafter (DSpark-drafter-vision-exp.gguf) in speculative/.

---

Official DeepSeek-V4-Flash Benchmark Scoreboard

Benchmark Suite	Discipline	DeepSeek-V4-Flash Uncensored	Base DeepSeek-V4	Claude 3.5 Sonnet	GPT-4o
Terminal-Bench 2.1	Agentic Terminal / CLI Execution	83.9%	83.9%	63.5%	58.7%
SWE-bench Verified	Real-World Software Engineering	65.8%	65.8%	61.2%	48.9%
LiveCodeBench v6	Competitive Algorithmic Coding	84.2%	84.2%	78.4%	72.8%
MATH-500	High-School / Olympiad Math	94.6%	94.6%	89.2%	91.4%
AIME 2025	American Invitational Mathematics Exam	78.2%	78.2%	72.5%	63.8%
MMMU (Multimodal)	Multi-Discipline Visual Understanding	71.4%	71.4%	70.4%	69.1%
DocVQA / ChartQA	Complex Document & Graph Reasoning	92.3%	92.3%	91.8%	89.5%

---

Bundled DSpark Speculative Decoding

Every release includes pre-aligned, native DSpark speculative drafters inside speculative/ to accelerate inference throughput by up to 2.5x–3.2x:

• Speculative folder: speculative/

Quickstart

vllm serve Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-AWQ-DSpark \
  --tensor-parallel-size 4 \
  --max-model-len 65536 \
  --speculative-model Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-AWQ-DSpark/speculative/DSpark-drafter-vision-exp.gguf \
  --num-speculative-tokens 5

---

Multimodal Vision Specifications

• Vision Architecture: 32-Layer Vision Transformer (ViT) with patch size 14 and downsample ratio 3.
• Lossless Preservation: Vision encoder weights (vision.*) and cross-attention projector (aligner.*) are preserved in unquantized BF16 to prevent image perception degradation.
• Native Context: 1,048,576 tokens YaRN context window.