Local Deployment of DeepSeek-V4.1-Flash Achieves High-Precision Performance on Legacy Hardware

A legal professional has reported the successful local deployment of the DeepSeek-V4.1-Flash model, achieving performance levels approaching NVIDIA’s NVFP4 precision using legacy V100 GPUs. The setup, which utilizes eight V100-SXM2-32GB graphics cards, aims to provide a private, enterprise-grade alternative to commercial frontier models like Claude.

According to a technical report shared on Reddit, the deployment—dubbed “Volta Max uPLIft”—utilizes a custom optimization strategy to manage memory and precision. The system maintains a 2-bit record of every routed expert while swapping the most frequently used experts to full NVFP4 precision. This method reportedly allows the hardware to maintain high precision for approximately 78% to 85% of decode routes.

Performance Benchmarks and Technical Specifications

The reported benchmarks, which were processed using Claude Code, indicate significant throughput for a local setup. The user reported the following metrics for the eight-card V100 configuration:

  • Decoding Speed: 118.9 tokens per second (tok/s) for single requests, increasing to 159.0 tok/s when handling two concurrent requests.
  • Context Window: The system supports a context of up to 524,288 tokens.
  • Precision Accuracy: The user claims the setup achieves the same next token as NVIDIA’s full NVFP4 pack on 95.1% of compared tokens, compared to 92.2% for a standard 2-bit build.
  • Prompt Processing: Prefill speeds for a 256K prompt were measured at approximately 2,329 to 2,430 tok/s.

The developer noted that while this high-precision configuration offers improved quality, it is slower in both decoding and prompt reading than a standard 2-bit build. The project, which includes custom kernels and optimizations, has been made public via GitHub under an Apache-2.0 license.

The deployment is designed to power a “legal harness,” a proprietary system of hooks, drafting engines, and RAG (Retrieval-Augmented Generation) stores. The user stated that the goal is to replicate the functionality of frontier models like Claude while ensuring that no data leaves the local environment.

The setup includes additional hardware, such as two DGX Sparks capable of running three to four concurrent DeepSeek sessions as agents with a 1M context window. The user described the result as a “production grade, workflow replacing local setup” intended for enterprise use, rather than a consumer-grade application.

Comparative Local Model Testing

In a separate technical demonstration, a user tested various coding agents—including Claude Code, OpenCode, GitHub Copilot Desktop, Pi, and Codex—against a local Qwen3.8 Flash Next model running on an AMD Ryzen AI Max+ 395 system.

The results, documented in a Reddit post, showed varying levels of efficiency across the different agents. Claude Code reached a final answer in approximately 14 minutes and 42 seconds, utilizing 31,373 output tokens. In contrast, Codex required over four hours to complete its task, generating 402,803 output tokens before the session was aborted. The test highlighted the significant variance in how different agent “harnesses” interact with local inference servers.