DeepSeek V4.1 Flash hands-on tests and comparisons

A blog post documents long-term use of a quantized uncensored DeepSeek V4.1 Flash 1017 model on two DGX Spark machines; a three-minute read.

A user ran the uncensored GSQ-RCO IQ2_XXS GGUF of DeepSeek-V4.1-Flash-1017 on a 24GB RTX Pro 4000 Blackwell, getting 36 tok/s decode and 400 tok/s prefill with working thinking, and plans to use it via DeepSeek Harness with KV pre-caching.

A user says they have started using a locally deployed DeepSeek-V4.1-Flash-1017 for real work and find it very comfortable, without citing specific tasks or numbers.

Argues production metrics beat benchmarks: Kimi K3.1 Pro scores 55 on the intelligence index vs Flash's 39–45, but Flash costs $0.48 per task vs $2.11, a 4.4x saving offset by heavier output token use.

After heavy batch usage, the user warns that input/output list prices mislead: cache hit rate and cache pricing make up a large share of the bill, and Flash 1017 stands out on both.

A prompt-driven DeepSeek Flash 1017 run produced a colorful voxel pagoda garden with drag-orbit, scroll zoom, auto tour and a time-of-day lighting slider, all in one HTML file.

Reading Qwen4 and DeepSeek together, the author notes Strata docs claim a 200GB+ MoE like V4.1 Flash can run on one RTX 3060 plus 46GB RAM at 8–10 tok/s, with under-40GB RAM untested.

The poster argues hacking is mostly diligence, and calls V4.1 Flash 1017 a free and underrated tool for offensive work, without concrete examples.

Across 8 live web apps, Flash 1017 scored F1 75.08 at $32.93, ahead of GLM-5.3's 73.53/$305, GPT-5.6 Sol's 69.80/$1,983 and Opus 5's 70.12/$3,032.

The 763B-parameter model ships in 1.75/2.00/2.25-bit quantizations totaling 239–274GB split across VRAM, RAM and NVMe; 2.00-bit raises perplexity 15.5%, and image input and parallel requests are disabled.

An HN post lists eight OpenAI and Anthropic price cuts in two months, including an 80% cut on GPT-5.6 Luna, and credits cheap open models such as DeepSeek V4.1 Flash as the pressure behind them.

vLLM v0.32.0 ships DaoCloud's H20 inference optimizations, reporting roughly 60% higher DeepSeek-V4.1 throughput.