Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Eight-V100 workflow brings DeepSeek-V4.1-Flash closer to NVFP4 precision

AI summary

A builder released the public Volta Max uPLIft workflow for running DeepSeek-V4.1-Flash on eight Nvidia V100 GPUs with a 524,288-token context. Its hybrid quantization keeps all routed experts in 2-bit form while storing the most frequently used experts at NVFP4 precision. In the builder’s tests, the build matched the next token produced by Nvidia’s full NVFP4 package on 95.1% of compared tokens, up from 92.2% for the 2-bit build. Code output slowed from 160–215 to 150–180 tokens per second. The builder also reports that Claude Code wrote the custom kernels and ran the benchmarks under his direction.

In-depth insight

The project makes its code, kernels, measurements and documented failures public, providing a detailed way to run a large-context DeepSeek model locally on eight older V100 GPUs. Its reported gains in reference fidelity come at a measured performance cost, while the code and AI-assisted development are the builder’s own work.

#DeepSeek#NVIDIA#Volta Max uPLIft#Claude Code#Claude Opus 5.5#Claude Fable 5.2#NVIDIA V100#NVIDIA NVFP4

Original · 3 min readRead the original
Group chat

No comments yet. Start the conversation.