DeepSeek V4.1 Flash hands-on tests and comparisons

A whatmodel.sh chart puts Chinese open models 7 months behind US closed ones as of October 2026, with DeepSeek V4.1 Flash as the latest open point and price left out of the comparison.

On a 100-vulnerability test Kimi K3.1 Pro found 47 for $124.69, while DeepSeek V4.1 Flash 1017 found 10 more and cut cost per find from $3.93 to $2.65.

Using DeepSeek V4.1 Flash in ZCode for small CSS fixes, the author saw 108 million tokens after 14–16 messages because the agent resends full history and never compacts; the same model feels normal in Claude Code.

The vLLM team reports DeepSeek V4.1 Flash on SemiAnalysis AgentX reaching 954K tokens/s/GPU total throughput and 196 tok/s P90 per user, worth $38.96/GPU-hour at DeepSeek off-peak pricing.

A link to llama.cpp pull request #30817, which integrates DeepSeek-V4.1-Flash into llama.cpp for local inference.

A user reports that the first LG SmartTV was rooted using DeepSeek V4.1 plus reverse-engineering tooling, arguing open models are now good enough for taking software apart.

DeepSeek's yifandotqiao thanked a third-party blog for serving V4.1 Flash with vLLM and reproducing their results, and acknowledged the DeepInfer Engram RDMA fetch bottleneck the blog identified.

A published recipe runs DeepSeek V4.1 Flash with TP4 on four DGX Sparks (300K daily, 1M max context) while each Spark also hosts a ComfyUI studio for Qwen Image 2.1 and MiniMax H3.

The author says their DeepSeek-V4.1-Flash benchmarks use real input distributions, realistic expert routing and the same serving configs validated on SemiAnalysis AgentX, and were online Oct 20 before DeepInfer shipped.

On three V100 32GB cards with dual 2650v2 and DDR3, the GSQ-RCO-IQ2_XXS quant decodes at 26.05 tok/s and prefills at 393.53 tok/s, versus 17.37 tok/s on two cards, with hardware-tuned builds clearly faster.

SGLang reports up to 1.04× DeepSeek's in-house prefill throughput on four B300/GB300 GPUs and up to 7.4% shorter decode steps on 64 GB300s, with vLLM matching DeepInfer's decode curves.

Working backward from DeepSeek's inference deployment Q&A figures on prefill/decode throughput and cache hit rate, the poster estimates one cluster at roughly 1,185 GPU pairs, or about 2,400 B200s.