Model Overview
Solstice-AI/DeepSeek-V4-Flash-Vision-UNCENSORED-GGUF-DSpark provides the official, high-efficiency GGUF quantization of DeepSeek-V4-Flash-Vision-UNCENSORED. This release preserves the unquantized BF16 multimodal vision encoder alongside MXFP4 quantized MoE weights, packaged with pre-aligned DSpark speculative drafters for low-latency inference.
Key Specifications
| Attribute | Specification |
|---|---|
| Base Model | orcarouter/DeepSeek-V4-Flash-Vision-Uncensored |
| Total Parameters | 305B (256 routed MoE experts, ~18B active per token) |
| Context Window | 1,048,576 tokens (1M YaRN native context) |
| Multimodal Vision | 32-layer Vision Transformer (ViT) in native BF16 (mmproj-BF16.gguf, 1.11 GB) |
| Base Model Quant | 4-shard GGUF MXFP4 (DeepSeek-V4-Flash-Vision-Uncensored-MXFP4-*.gguf) |
| Speculative Drafter | Bundled DSpark semi-autoregressive drafter (speculative/DSpark-drafter-vision-exp.gguf, 6.94 GB) |
Benchmark Highlights
- Terminal-Bench 2.1: 83.9% (Agentic CLI execution)
- SWE-bench Verified: 65.8% (Real-world software engineering)
- LiveCodeBench v6: 84.2% (Algorithmic problem solving)
- MATH-500: 94.6%
- DocVQA / ChartQA: 92.3% (Complex visual reasoning & document grounding)
Attribution & Acknowledgments
- Original Architecture: DeepSeek AI
- Uncensored Alignment: orcarouter
- Quantization & DSpark Packaging: Solstice-AI