Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark overview

---

pipeline_tag: image-text-to-text

base_model:

• deepseek-ai/DeepSeek-V4-Flash-Vision-Exp

license: mit

library_name: gguf

tags:

• gguf
• deepseek
• deepseek-v4
• vision
• multimodal
• image-text-to-text
• solstice-ai
• llama.cpp
• ollama
• dspark
• speculative-decoding
• draft-model
• ds4
• mmproj

---

<p align="center">

<img src="https://cdn-uploads.huggingface.co/production/uploads/67c2e844e0921a5410eec10a/Y5M42dCag2f7Fc6fDtV0Z.jpeg" alt="Solstice-AI Banner" width="100%">

</p>

<h1 align="center">DeepSeek-V4-Flash-Vision-Exp (Ultra-Optimized GGUF Suite)</h1>

<h3 align="center">Official Solstice-AI GGUF Release &bull; Unsloth Dynamic v3.0 Quants &bull; Lossless BF16 Vision Projector &bull; Bundled DSpark Drafter &bull; 1M Context</h3>

<p align="center">

<img src="https://img.shields.io/badge/org-Solstice--AI-blueviolet" alt="Solstice-AI">

<img src="https://img.shields.io/badge/license-MIT-blue" alt="License">

<img src="https://img.shields.io/badge/format-GGUF-brightgreen" alt="Format">

<img src="https://img.shields.io/badge/pipeline-image--text--to--text-success" alt="Pipeline">

<img src="https://img.shields.io/badge/context-1M%20Tokens-purple" alt="Context">

<img src="https://img.shields.io/badge/speculative-DSpark%20Drafter-red" alt="DSpark">

<img src="https://img.shields.io/badge/vision-BF16%20Projector-orange" alt="Vision">

</p>

---

Model Overview

Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark provides the curated, production-grade GGUF suite for DeepSeek's experimental multimodal foundation model deepseek-ai/DeepSeek-V4-Flash-Vision-Exp (305B total parameters, 256 routed MoE experts, 13B activated per token, 43 layers, Multi-Head Latent Attention).

This release bundles the complete multimodal stack: curated base quantizations, a standalone lossless BF16 multimodal vision projector, and the pre-aligned DSpark speculative drafter.

Artifact Matrix

| Artifact | Format | Shards | Size | Target Hardware & Role |

| :--- | :---: | :---: | :---: | :--- |

| UD-IQ4_XS | IQ4_XS (imatrix) | 4 | 136.7 GB | Recommended. Production sweet spot. Fits Apple M2/M3/M4 Ultra (192GB) or dual 80GB GPUs. |

| UD-Q3_K_XL | Q3_K_XL | 4 | 128.2 GB | High-Throughput Linear K-Quant. 15–20% faster raw dequantization. Fits 128GB–192GB setups. |

| UD-Q8_K_XL | Q8_K_XL | 5 | 161.9 GB | Bit-Exact Reference. Full precision across dense projections and router heads. |

| mmproj-BF16.gguf | BF16 | 1 | 934.5 MB | Multimodal Vision Projector. Kept in native BF16 to eliminate visual hallucinations. |

| speculative/DSpark-drafter-vision-exp.gguf | Q2_K / Q8_0 | 1 | 6.94 GB | DSpark Semi-Autoregressive Drafter. Tuned for Vision-Exp revision e46e16bf. Drives 1.4×–1.7× speculative acceleration in ds4. |

---

Architectural Breakdown & Specifications

• Total Parameters: 305B total, 13B active per token across 256 routed MoE experts (6 active per token).
• Attention Mechanism: Compressed Sparse Attention (CSA) and Multi-Head Latent Attention (MLA).
• Multimodal Vision Encoder: 32-layer Vision Transformer (ViT) with patch size 14 and downsample ratio 3.
• Context Window: 1,048,576 tokens (1M native YaRN context).

---

Citations & Acknowledgments

• Original Architecture & Weights: DeepSeek AI
• Quantization: Unsloth Dynamic v3.0
• DSpark Drafter Packaging: bleysg & Solstice-AI