Instructions to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS
# Run inference directly in the terminal:
llama cli -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS
# Run inference directly in the terminal:
llama cli -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS
# Run inference directly in the terminal:
./llama-cli -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Use Docker

docker model run hf.co/Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

LM Studio
Jan

vLLM

How to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Ollama
How to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with Ollama:
```
ollama run hf.co/Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS
```
Unsloth Desktop

How to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with Docker Model Runner:
```
docker model run hf.co/Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS
```

Lemonade

How to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Run and chat with the model

lemonade run user.DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark-UD-IQ4_XS

List all available models

lemonade list

Hermes Agent

How to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark:UD-IQ4_XS" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Solstice-AI Banner

DeepSeek-V4-Flash-Vision-Exp (Ultra-Optimized GGUF)

Official Solstice-AI Curated GGUF Release • Native DSpark Speculative Decoding • Pure W4A16 & Near-Lossless MoE Tiers • llama.cpp, Ollama & ds4

Original Model by DeepSeek-AI • Quantized with Unsloth Dynamic v3.0 • DSpark Drafter by bleysg • Curated, Verified & Packaged by Solstice-AI

Executive Summary

Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF provides the definitive, production-curated all-in-one GGUF stack for DeepSeek's experimental multimodal foundation model deepseek-ai/DeepSeek-V4-Flash-Vision-Exp.

DeepSeek-V4-Flash-Vision-Exp features a 305B parameter Mixture-of-Experts (MoE) architecture (13B activated parameters per token) with 43 hidden layers, 256 routed experts (6 active per token), Multi-Head Latent Attention (MLA), and a dedicated 32-layer multimodal vision encoder.

The All-in-One Speculative Ecosystem

Unlike disjointed repositories where users must track down separate drafter repositories and projector files, this release bundles the full production triangle:

Curated Base Models: Unsloth Dynamic v3.0 quants covering the optimal sweet spots (UD-Q8_K_XL, UD-IQ4_XS, UD-Q3_K_XL).
Lossless Vision Projector: mmproj-BF16.gguf preserving high-resolution OCR, coordinate grounding, and chart comprehension.
Native DSpark Speculative Drafter: speculative/DSpark-drafter-vision-exp.gguf accelerating inference by 1.4× to 1.7× (up to 2×+ on structured code/JSON) completely losslessly.

Artifact Catalog & Hardware Requirements

Category	Artifact	Format	Shards	Size	Min RAM/VRAM	Role & Hardware Target
Base Model	`UD-IQ4_XS`	IQ4_XS (imatrix)	4	136.7 GB	160 GB	Recommended. Production sweet spot. Saves 25.2 GB off Q8; >99% SWE-bench accuracy. Fits Apple M2/M3/M4 Ultra (192GB) or dual 80GB GPUs.
Base Model	`UD-Q3_K_XL`	Q3_K_XL	4	128.2 GB	144 GB	High-Throughput Linear K-Quant. 15–20% faster raw dequantization over IQ formats. Fits 128GB–192GB memory setups.
Base Model	`UD-Q8_K_XL`	Q8_K_XL	5	161.9 GB	192 GB	100% Bit-Exact Reference. Full precision across dense projections and router heads. For evaluation and enterprise rigor.
Vision	`mmproj-BF16.gguf`	BF16	1	934.5 MB	+1.5 GB	High-Precision Multimodal Projector. Kept in native BF16 to eliminate visual hallucination and coordinate drift.
Acceleration	`speculative/DSpark-drafter-vision-exp.gguf`	Q2_K / Q8_0	1	6.94 GB	+7.0 GB	DSpark Semi-Autoregressive Drafter. Tuned specifically for Vision-Exp revision `e46e16bf`. Drives 1.4×–1.7× lossless speculative decoding.

What is DSpark Speculative Decoding?

DSpark is DeepSeek's official speculative decoding framework (from the DeepSpec architecture).

Traditional speculative decoding (pure parallel drafting) suffers from severe acceptance decay: while token 1 has a high acceptance rate, tokens 4 through 7 are almost always rejected because parallel drafters cannot model inter-token dependencies.

How DSpark Solves This:

Semi-Autoregressive Drafter: The drafter combines 3 full DeepSeek transformer layers (attention + 256-expert MoE) with a rank-256 Markov head (markov_w1/markov_w2). Draft tokens are conditioned on previous draft tokens inside the window.
Deep Target Fusion: Fuses activation representations across the target model's final layers (layers 40, 41, and 42) via dedicated main_proj and main_norm matrices.
High Acceptance Rate: Achieves 52% to 91% acceptance rates, committing a mean of ~3.7 tokens per verification step.
100% Bit-Exact & Lossless: The target 305B model verifies every single drafted token. Output is mathematically identical to running the base model without speculative decoding.

Quick Start Guide

1. Download the Complete Stack via Hugging Face CLI

You can download the recommended UD-IQ4_XS base model, the BF16 vision projector, and the DSpark drafter in one command:

hf download Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF \
  --include "UD-IQ4_XS/*" "mmproj-BF16.gguf" "speculative/DSpark-drafter-vision-exp.gguf" \
  --local-dir ./DeepSeek-V4-Flash-Vision-Exp-GGUF

2. Serving with `llama-server` + DSpark Acceleration

Launch an OpenAI-compatible API server with multimodal vision support and DSpark speculative acceleration:

./llama-server \
  -m ./DeepSeek-V4-Flash-Vision-Exp-GGUF/UD-IQ4_XS/DeepSeek-V4-Flash-Vision-Exp-UD-IQ4_XS-00001-of-00004.gguf \
  --mmproj ./DeepSeek-V4-Flash-Vision-Exp-GGUF/mmproj-BF16.gguf \
  --model-draft ./DeepSeek-V4-Flash-Vision-Exp-GGUF/speculative/DSpark-drafter-vision-exp.gguf \
  --spec-draft-n-max 7 \
  -c 32768 \
  -ngl 99 \
  --host 0.0.0.0 \
  --port 8080

llama.cpp automatically loads and sequences all 4 sequential shard files (00001-of-00004 through 00004-of-00004). You only need to pass the path to the first shard.

Depending on your llama.cpp build version, the speculative drafting flags are --model-draft (or --spec-draft-model) and --spec-draft-n-max (older builds used --draft-model and --draft-max).

3. CLI Interactive Inference with Vision & DSpark

./llama-cli \
  -m ./DeepSeek-V4-Flash-Vision-Exp-GGUF/UD-IQ4_XS/DeepSeek-V4-Flash-Vision-Exp-UD-IQ4_XS-00001-of-00004.gguf \
  --mmproj ./DeepSeek-V4-Flash-Vision-Exp-GGUF/mmproj-BF16.gguf \
  --model-draft ./DeepSeek-V4-Flash-Vision-Exp-GGUF/speculative/DSpark-drafter-vision-exp.gguf \
  --spec-draft-n-max 7 \
  -p "Describe the architecture diagram and extract all code snippets from the screenshot:" \
  --image ./diagram.png \
  -c 16384 \
  -ngl 99

4. Serving with `ds4` on NVIDIA DGX Spark / CUDA Workstations

If serving on an NVIDIA DGX Spark (GB10 Grace Blackwell) or CUDA GPU server using Entrpi/ds4:

env DS4_CONT_DSPARK=1 \
    DS4_DSPARK_MODEL=./DeepSeek-V4-Flash-Vision-Exp-GGUF/speculative/DSpark-drafter-vision-exp.gguf \
  ./ds4-server --cuda \
  -m ./DeepSeek-V4-Flash-Vision-Exp-GGUF/UD-IQ4_XS/DeepSeek-V4-Flash-Vision-Exp-UD-IQ4_XS-00001-of-00004.gguf \
  -c 32768 \
  --port 8000

Architectural Breakdown & Pruning Rationale

                 +-------------------------------------------------------------+
                 |          DeepSeek-V4-Flash-Vision-Exp (305B Total)           |
                 +-------------------------------------------------------------+
                                                |
               +--------------------------------+-------------------------------+
               | (~96% of Total Parameters)                                     | (~4% of Parameters)
               v                                                                v
+-------------------------------+                              +-------------------------------+
|  256 Routed MoE Experts       |                              |  Non-Expert Dense Layers      |
|  (Pre-trained in OCP MXFP4)   |                              |  (MLA, Attention, Routers,    |
+-------------------------------+                              |   Shared Experts, Embeddings) |
               |                                               +-------------------------------+
               |                                                                |
               v                                                                v
   Native ~4.25-bit Base State                                       Native BF16 / FP8 / FP32

Why Sub-3-Bit Tiers Were Pruned

Native 4-Bit Base: DeepSeek pre-trained all 256 routed MoE experts in OCP MXFP4. The experts already occupy ~4.25 bits natively.
Redundant Q4: Downcasting the dense layers to Q4 saves only ~6.8 GB compared to Q8 while injecting noise into router selection.
Sub-3-Bit Quality Cliff: Downcasting MoE experts below 3 bits (Q2, IQ2, IQ1) causes catastrophic router misallocation—SWE-bench drops from ~90% down to ~70% (worse than a 27B dense model).
The Optimal Matrix: By curating UD-Q8_K_XL, UD-IQ4_XS, UD-Q3_K_XL, and the standalone DSpark drafter, Solstice-AI delivers maximum fidelity, minimal VRAM waste, and industry-leading inference throughput.

Repository File Tree & Checksums

UD-IQ4_XS/
├── DeepSeek-V4-Flash-Vision-Exp-UD-IQ4_XS-00001-of-00004.gguf (5.3 MB)
├── DeepSeek-V4-Flash-Vision-Exp-UD-IQ4_XS-00002-of-00004.gguf (49.4 GB)
├── DeepSeek-V4-Flash-Vision-Exp-UD-IQ4_XS-00003-of-00004.gguf (49.6 GB)
└── DeepSeek-V4-Flash-Vision-Exp-UD-IQ4_XS-00004-of-00004.gguf (37.6 GB)

UD-Q3_K_XL/
├── DeepSeek-V4-Flash-Vision-Exp-UD-Q3_K_XL-00001-of-00004.gguf (5.3 MB)
├── DeepSeek-V4-Flash-Vision-Exp-UD-Q3_K_XL-00002-of-00004.gguf (49.4 GB)
├── DeepSeek-V4-Flash-Vision-Exp-UD-Q3_K_XL-00003-of-00004.gguf (49.5 GB)
└── DeepSeek-V4-Flash-Vision-Exp-UD-Q3_K_XL-00004-of-00004.gguf (29.3 GB)

UD-Q8_K_XL/
├── DeepSeek-V4-Flash-Vision-Exp-UD-Q8_K_XL-00001-of-00005.gguf (5.3 MB)
├── DeepSeek-V4-Flash-Vision-Exp-UD-Q8_K_XL-00002-of-00005.gguf (49.2 GB)
├── DeepSeek-V4-Flash-Vision-Exp-UD-Q8_K_XL-00003-of-00005.gguf (49.7 GB)
├── DeepSeek-V4-Flash-Vision-Exp-UD-Q8_K_XL-00004-of-00005.gguf (49.5 GB)
└── DeepSeek-V4-Flash-Vision-Exp-UD-Q8_K_XL-00005-of-00005.gguf (13.5 GB)

Vision Projector:
└── mmproj-BF16.gguf (934.5 MB)

Speculative Acceleration:
└── speculative/DSpark-drafter-vision-exp.gguf (6.94 GB)

Citations & Acknowledgments

@misc{deepseek2026v4flashvision,
  title={DeepSeek-V4-Flash-Vision-Exp: An Open Experimental Multimodal MoE Architecture},
  author={DeepSeek-AI Team},
  year={2026}
}

@software{deepseek2026deepspec,
  title={DeepSpec: High-Throughput Semi-Autoregressive Speculative Decoding Framework},
  author={DeepSeek-AI Team},
  year={2026},
  url={https://github.com/deepseek-ai/DeepSpec}
}

@software{unsloth2026dynamic,
  title={Unsloth Dynamic Quantization v3.0},
  author={Unsloth AI Team},
  year={2026}
}

@software{bleysg2026dspark,
  title={DSpark Block Drafter for DeepSeek-V4-Flash-Vision-Exp},
  author={bleysg},
  year={2026},
  url={https://huggingface.co/bleysg/DeepSeek-V4-Flash-DSpark-drafter-GGUF}
}

@software{solsticeai2026gguf,
  title={Solstice-AI Ultra-Optimized DeepSeek-V4 Curation & Speculative Matrix},
  author={Solstice-AI Research Team},
  year={2026},
  url={https://huggingface.co/Solstice-AI}
}

Downloads last month: 317

GGUF

Model size

284B params

Architecture

deepseek4

Hardware compatibility

3-bit

4-bit

8-bit

View +1 variant

Model tree for Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark

Base model

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

Quantized

(24)

this model

DeepSeek-V4-Flash-Vision-Exp (Ultra-Optimized GGUF)

Official Solstice-AI Curated GGUF Release • Native DSpark Speculative Decoding • Pure W4A16 & Near-Lossless MoE Tiers • llama.cpp, Ollama & ds4

Executive Summary

The All-in-One Speculative Ecosystem

Artifact Catalog & Hardware Requirements

What is DSpark Speculative Decoding?

How DSpark Solves This:

Quick Start Guide

1. Download the Complete Stack via Hugging Face CLI

2. Serving with llama-server + DSpark Acceleration

3. CLI Interactive Inference with Vision & DSpark

4. Serving with ds4 on NVIDIA DGX Spark / CUDA Workstations

Architectural Breakdown & Pruning Rationale

Why Sub-3-Bit Tiers Were Pruned

Repository File Tree & Checksums

Citations & Acknowledgments

Model tree for Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-UltraOptimised-GGUF-DSpark

2. Serving with `llama-server` + DSpark Acceleration

4. Serving with `ds4` on NVIDIA DGX Spark / CUDA Workstations