Intel / post

vLLM Runs DeepSeek v4.1 Flash Day-0 on NVIDIA, Stalls on AMD

Source: x.com
Date: Published Sep 11, 2026

From the Daily Brief

SemiAnalysis reports that on DeepSeek V4.1 Flash's release day, NVIDIA's vLLM stack ran the model cleanly across hardware from H100 through the newer GB300, while AMD's announced ROCm image for the same model had still not shipped a full day after release. Day-zero framework support is what decides whether a new open model can actually be deployed in production the day it launches, rather than sitting unusable while teams wait for a serving stack to catch up. This gap follows the same pattern seen with earlier open model releases, where NVIDIA's software ecosystem consistently reaches parity with a new release faster than AMD's does. For teams evaluating AMD hardware as a cost alternative for serving open models, this is a concrete, dated data point on how far that software gap still runs, not a one-off complaint.

Read the 2026-09-13 Brief →

Context

vLLM is a widely used open-source system for serving large language models efficiently. On the day DeepSeek released its v4.1 Flash model, SemiAnalysis reports that NVIDIA's vLLM support worked with zero issues across all six of its current GPU types: H100, H200, B200, B300, GB200 and GB300.

By contrast, SemiAnalysis reports AMD's vLLM support for the same model was still not working as of that post. Two days later, AMD did release its own image for the model, but SemiAnalysis found it cost far more per dollar to run than NVIDIA's hardware once it arrived.

Read the source x.com