Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-NVFP4

DeepSeek-V4-Flash-Vision-Exp (NVFP4 W4A16)

Official Solstice-AI NVFP4 Release • Lossless MXFP4-to-NVFP4 MoE Transcode • Native Blackwell & vLLM Acceleration

---

Model Overview

Solstice-AI/DeepSeek-V4-Flash-Vision-Exp-NVFP4 is the high-performance NVFP4 (W4A16) derivative of DeepSeek's experimental multimodal foundation model deepseek-ai/DeepSeek-V4-Flash-Vision-Exp, optimized specifically for NVIDIA Blackwell (B200 / GB200 / DGX Spark) and modern inference engines like vLLM and TensorRT-LLM.

Features NVIDIA TensorRT ModelOpt's bit-exact MXFP4-to-NVFP4 transcode, mapping all 11,008 expert linear matrices into NVFP4 block-scale format with zero precision loss while preserving the 32-layer vision tower, attention heads, router gates, and shared experts in native high precision.

Key Specifications

Attribute	Specification
Base Model	deepseek-ai/DeepSeek-V4-Flash-Vision-Exp
Total Parameters	305B (256 routed MoE experts, 13B activated per token, 43 layers)
Context Window	1,048,576 tokens (1M native YaRN context)
Multimodal Vision	32-layer Vision Transformer (ViT) in native BF16 (vision.*, aligner.*)
Quantization Format	SafeTensors NVFP4 (W4A4 MoE experts, group size 16, block-scale calibration)
Target Hardware	NVIDIA Blackwell (B200 / GB200 / DGX Spark / RTX 5090)

---

Benchmark Highlights

Benchmark Suite	Focus	Baseline Checkpoint	Solstice-AI NVFP4	Delta
GSM8K (8-shot)	Multi-Step Mathematical Reasoning	97.0%	97.0%	0.0 pt (Zero Loss)
Long-Context Retrieval	32K → 250K Tokens Needle Test	100%	100% (12/12 gates)	0.0% Drift
Multimodal Vision QA	Fine-Grained Object & Text Recognition	Passed	Passed (Exact Match)	Sub-50ms TTFT

---

License & Attribution

• Model License: MIT License (inherited from deepseek-ai/DeepSeek-V4-Flash-Vision-Exp).
• Quantization Pipeline: Produced using NVIDIA TensorRT ModelOpt; original recipe by s-zaizen.
• Packaging & Curation: Maintained by Solstice-AI.