Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MXFP4-DSpark

DeepSeek-V4-Flash-Vision-Exp (Native MXFP4 • W4A16)

Official Solstice-AI MXFP4 Production Release • Native Pre-trained OCP Microscaling FP4 • Pure W4A16 Execution in vLLM & SGLang

Original Architecture by DeepSeek-AI • Curated, Verified & Packaged by Solstice-AI

---

Executive Summary

Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MXFP4 is the canonical, production-optimized OCP MXFP4 (W4A16) distribution of 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, natively supporting a 1,048,576-token (1 Million) context window.

The Gold Standard for W4A16 GPU Serving

Unlike models subjected to lossy post-training weight quantization, DeepSeek pre-trained all 256 routed MoE experts natively in Open Compute Project (OCP) MXFP4 block-floating-point format.

When deployed on modern GPU clusters (NVIDIA H100, A100, RTX 4090, or Blackwell) via engines like vLLM and SGLang, this checkpoint runs strictly in W4A16 mode:

• Weights in 4-Bit (OCP MXFP4): Dramatically reduces memory bandwidth pressure and fits comfortably on dual or quad 80GB GPU nodes (~150 GB total checkpoint footprint).
• Activations in 16-Bit (BF16): Activations remain uncompressed during matrix multiplication, ensuring zero activation outlier clipping, zero precision drift, and 100% mathematical integrity across long-context reasoning and vision grounding.
• Preserved High-Precision Non-Expert Stack: All 32 layers of the multimodal vision encoder, router gate logits, attention heads, shared experts, and layer normalizations remain in their native BF16 / FP8 precision.

---

Architectural Breakdown

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

---

Hardware & Deployment Matrix

GPU Configuration	Total VRAM	Context Window	KV Cache Precision	Deployment Mode
2× NVIDIA A100 / H100 80GB	160 GB	32,768	FP8	High-throughput API server (compact context)
4× NVIDIA A100 / H100 80GB	320 GB	131,072	FP8 / BF16	Recommended. Production long-context agent serving
8× NVIDIA H100 / H200 80/141GB	640+ GB	1,048,576 (1M)	FP8 / BF16	Maximum concurrency & full 1M reasoning window

---

Official Benchmark Scoreboard

Evaluated by DeepSeek using the minimal mode of DeepSeek Harness with max reasoning effort (temperature = 1.0, top_p = 0.95):

Benchmark Suite	Focus Area	DeepSeek-V4-Flash-Vision-Exp	DeepSeek-V4-Flash-0731	Claude Opus-4.8
Terminal Bench 2.1	Terminal Execution & Shell Control	83.9	82.7	85.0
NL2Repo	Code Repository Generation	57.7	54.2	69.7
Cybergym	Cybersecurity Operations & CTF	75.3	76.7	78.3
DeepSWE	Real-World Software Engineering	59.3	54.4	58.0
Toolathlon-Verified	Multi-Tool Function Calling Loops	75.9	70.3	76.2
DSBench-Hard	Complex Data Science Workflows	63.6	59.6	71.7
AutomationBench (Public)	End-to-End Task Automation	25.7	25.1	27.2
ApexBench (Pass@1)	Multimodal Agent UI & Vision QA	36.5	26.2	39.4
Agents' Last Exam	Multimodal Reasoning & Academic VQA	27.3	25.2	25.7
Chartography	High-Density Chart & Visual Analysis	64.3	—	65.0
ZeroBench (Pass@5)	Zero-Shot Multimodal Complex Tasks	35.0	—	34.0

---

Quick Start & Serving Guide

1. High-Throughput Serving with vLLM + Native DSpark Speculative Acceleration

Launch an OpenAI-compatible API server with multimodal vision support, native reasoning parser, tool calling, and DSpark speculative decoding:

vllm serve Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MXFP4 \
  --tensor-parallel-size 4 \
  --kv-cache-dtype fp8 \
  --block-size 256 \
  --max-model-len 131072 \
  --tool-call-parser deepseek_v4 \
  --enable-auto-tool-choice \
  --reasoning-parser deepseek_v4 \
  --speculative-config '{"method":"dspark","num_speculative_tokens":3,"draft_sample_method":"probabilistic","enable_adaptive_verification":true}' \
  --trust-remote-code \
  --port 8000

Target and draft weights are co-located in the same checkpoint (shards 46–48). You do not need to supply an external draft model path.

---

2. Serving with SGLang

For maximum throughput and low-latency prefix caching in agentic workflows, enable DSpark natively:

sglang serve \
  --model-path Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MXFP4 \
  --tp 4 \
  --speculative-algorithm DSPARK \
  --mem-fraction-static 0.85 \
  --trust-remote-code \
  --host 0.0.0.0 \
  --port 30000

---

4. Interactive Multimodal Python Client Example

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="none")

response = client.chat.completions.create(
    model="Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-MXFP4",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Analyze this technical architecture diagram and extract all component interactions:"},
                {
                    "type": "image_url",
                    "image_url": {
                        "url": "https://example.com/system_architecture.png"
                    }
                }
            ]
        }
    ],
    max_tokens=2048,
    temperature=0.2
)

print(response.choices[0].message.content)

---

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{solsticeai2026mxfp4,
  title={Solstice-AI DeepSeek-V4-Flash-Vision-Exp MXFP4 Distribution},
  author={Solstice-AI Research Team},
  year={2026},
  url={https://huggingface.co/Solstice-AI}
}