DeepSeek V4.1 Flash hands-on tests and comparisons

Testing ISTA-DASLab's IQ2-XXS GGUF on a 32GB Pro4500 and a 16GB 5080 build, the author finds VRAM minus shared weights decides how many experts fit, and more DRAM helps less than more VRAM.

A short scenario-based take: DeepSeek V4.1 Flash is fast and cheap to get started, Kimi K3.1 Pro handles long jobs if you ration quota, and Claude Opus 5.5 is pricey but avoids rework.

A hands-on test has ds-v4.1-flash build a procedural steam-harbour 3D open world with PBR materials, SSAO, SMAA and a day/night cycle all in code, using 152M tokens at 99.8% cache hit.

A user reports running DeepSeek V4.1 Flash for free on OpenCode Zen at roughly 110 tok/s, with a screenshot attached.

On the black-box WebExploitBench offensive-security benchmark, Kimi-K3.1-Pro leads at 46.43%, DeepSeek-V4.1-Flash-1017 is second at 42.86%, and GLM-5.3-Flash third at 35.71%.

On Mazebench, Kimi K3.1 Pro is the first open-weights model with a nontrivial score at 2%, while DeepSeek V4.1 Flash follows at 1%.

A user runs V4.1 Flash on an ordinary 64GB-RAM dev desktop using Strata plus gsq-rco quants, arguing that shrinking open weights make per-token mid-tier pricing harder to justify.

The author suggests letting smarter Kimi K3.1 Pro do the reasoning while fast, cheap V4.1 Flash 1017 handles output, asking how hard that split would be.

A 367-task eval of official V4.1 Flash 1017 against a local GSQ-RCO IQ2_XS quant found both passing everything, with the full-weight model scoring marginally lower than the quantized one.

The model card lists Terminal-Bench 4.0 rising from 31.2 on the previous Flash to 42.3, plus deepseek-recipe Rust libraries with Python bindings.

The author released a Maya-S GGUF quant built on DeepSeek-V4.1-Flash-1017, running about 20 tok/s on an RTX 5090 with quality they call impressive.

A home rig with two RTX 3090s and 256GB DDR4 runs DeepSeek V4.1 Flash at EXL3 ~3bpw, hitting 26–32 tok/s decode and 737–830 tok/s prefill, with a 100k prompt in about 2 minutes.