dgx-spark-deepseek-v41

High-performance DeepSeek-V4.1-Flash implementation optimized for DGX Spark systems.

PlayableOpen source

Description

An optimized quantization of DeepSeek-V4.1-Flash using EXL3 trellis tensors specifically tuned for tensor parallel serving on NVIDIA DGX Spark hardware. The project includes custom vLLM patches and a specialized MoE plugin to achieve over 100 tokens per second while maintaining high accuracy across benchmark suites.

Descriptions, tags, and model credits may be AI-generated or inferred from public sources and can be incomplete or wrong. Learn more.

Details

Why this is interesting: Impressive engineering to optimize DeepSeek-V4.1-Flash for specialized hardware with custom EXL3 experts and vLLM…
Attributed to: @sudoingx
Built with: V4.1 Flash
Links: huggingface.co
View tweet
View GitHub repo
Tags: #quantization #vllm #dgx-spark #deepseek

Discovered: X
First posted: X · Oct 14, 2026
Listed: Added 1h ago
Build: Strong
Models: Strong