DeepSeek V4.1 Flash hands-on tests and comparisons

oMLX 0.7.2 raises DeepSeek V4.1 decode from 49.4 to 68.2 tok/s on the oQ3e quant, with prefill holding near 900 tok/s up to a 128K prompt on an M3 Ultra 512GB.

A community evaluation in DeepSeek Harness rates DeepSeek V4.1 Flash 1017 as roughly comparable to GPT-5.6 Sol at xhigh effort.

A post quotes a claim that DeepSeek V4.1 Flash 1017 beat GPT 6 Sol on multi-step offensive cybersecurity tasks, stressing it is open and locally runnable.

A hands-on test has DeepSeek V4.1 Flash generate a rainy neon Tokyo street scene in a single HTML file with no images or libraries, taking 4.5 minutes and $0.05 with one fix needed; the full prompt is included.

A user testing the non-abliterated GGUF says tool calling in the llama.cpp PR branch requires --jinja plus the DeepSeek-V4.1-Flash chat template file.

A user reports Unsloth's llama.cpp build now supports DeepSeek V4.1 Flash with almost no drop in inference speed.

vLLM with Wide-EP and EPLB on 32 GB200s reaches over 94% of DeepInfer's 131K-context decode throughput for V4.1 Flash, beats it by 3.6% at bs256, and breaks down per-GPU cost and break-even utilization.

Local deployment with BLACKFOLD on two DGX Sparks reports 2,008 tok/s prefill at 128K context and 113 tok/s code decode, with a rerun planned.

Running DeepSeek V4.1 Flash locally via a llama.cpp PR started at 27 tok/s and dropped to 11 tok/s at 50k tokens; after a fix it only falls to 24 tok/s, though agent use still fails, so the author will try Unsloth next.

A hands-on test runs a quantized V4.1 Flash on a 48GB modded 4090 with 128GB RAM, caching hot experts in VRAM to lift throughput from 5–6 to nearly 20 tok/s.

An analysis of DeepSeek open-sourcing its production inference stack DeepInfer, which lists Huawei Ascend 950 alongside Nvidia SM100 and aims to shift the default optimization target away from CUDA.

A user pushes back on local-deployment advice: a $3–5k machine only runs quantized 27–32B models, while DeepSeek V4.1 Flash needs 4x RTX PRO 6000 or 4x DGX Spark, so frontier API subscriptions remain the realistic path.