DeepSeek V4.1 Flash hands-on tests and comparisons

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.

The author says GPU rental prices are surging and suggests buying used 30/40-series cards on Xianyu, unlocking cores with strata, and deploying DeepSeek V4.1 Flash, without providing any benchmark data.

On FrontierCode's maintainer-would-merge grading, Opus 5.5 in Claude Code scores 56.0% versus 52.6% for Codex with DeepSeek V4.1 Flash, at $11.42 vs $0.39 per task.

Four days after DeepSeek open-sourced its inference engine, vLLM claims parity with DeepInfer on V4.1-Flash decode and SGLang claims up to 7.4% faster with a repro guide.

The author plugged V4.1 Flash into the Claude Code terminal and spent 5.5 hours updating npm packages, project deps, WSL, Windows and a VPS OS for $0.45 total.

Same voxel-world prompt: two Sparks (EXL3 2.5bpw) took 73.7 min at ~70 tok/s decode with a 5.5/10 self-rating; four Sparks (4.0bpw) took 49.7 min at ~82 tok/s with 7.5/10, though differing quant and recipe versions prevent isolating Spark count.