DeepSeek V4.1 Flash hands-on tests and comparisons

Official DeepSWE v1.1 numbers across five harnesses range from mini-SWE at 74.2% to OpenCode at 65.5%, an 8.7-point spread for the same model.

A comment notes that an engine like Strata can run a DeepSeek V4.1 quant sized for 256GB of memory by offloading portions, leaving some system memory after loading.

Same prompt in DeepSeek Harness compares Mia's FP8/MXFP4 checkpoint with bertholomus' EXL3 4.0bpw quant: 93.5 min and 258,931 tokens for an 8.5/10 world versus 49.7 min and 121,707 tokens for 7.5/10, at 82.8 vs 69.5 tok/s single-request.

On a dual Tesla V100 32GB, 96GB RAM server, swapping in a Ryzen 9 5900XT pushed the 340GB Maya-S quant of V4.1 Flash past 20 tok/s decode.

On self-hosted vLLM, Hermes and Claude Code both hit 75% and DSH 70%; DSH uses 1.23x Hermes prompt tokens while Claude Code uses 0.74x but caches only 9–11% of prefixes.

DeepInfer's open offline-inference release bundles FlashMLA, DeepGEMM, DeepEP and DSpark decoding, runs on NVIDIA SM100 and Ascend 950, and is Apache 2.0 licensed.

Nine days after launch, DeepSeek V4.1 Flash 1017 used 109.4B tokens on LLM Gateway, 34.6% of the weekly total and up 340.3%, while the original V4.1 Flash fell 45.6% to 86.7B.

On four DGX Sparks in TP4, V4.1 Flash 1017 generated an 875K-triangle floating-isle world in one Harness run, scoring 9/10 in the browser check at about 70 tok/s decode.

Codex blocked the author's chat over a cybersecurity risk while testing a local security model, so they moved that work to DeepSeek V4.1 Flash, arguing refusals justify keeping more than one model.

A heavy user says billions of tokens over weeks on DeepSeek V4.1 Flash 1017 never triggered a safety refusal, and asks what others are prompting.

DeepSeek V4.1 Flash ties GPT-6.1 Sol at 96.0 on real merged-PR bug fixes at $0.18 per successful task (38x cheaper than Fable), but scores 40.9 vs 77.9 on long agentic work.

A user compares V4.1 Flash with GPT-6.1 Sol high and argues the 20% price gap is offset by more rework, so GPT may cost less overall; Flash excels at office output and multi-turn delivery, with pure reasoning between GPT-5.5 high and GPT-5.6 Sol full.