DeepSeek V4.1 Flash hands-on tests and comparisons

A blogger's DGX Spark recommendations: one Spark runs Qwen3.8-Flash-Next, while two or more can run DeepSeek-V4.1-Flash-1017 (stronger on some tasks) or GLM-5.3-Flash (more consistent), with both performing well on four Sparks.

A user serving DeepSeek V4.1 Flash 1017 with ViC305 quantization and vllm-exl3 got it loading stably after a few hours, but throughput is just 1 tok/s; root cause still being investigated.

int21_ai says its generated Rust Inference Factory beats vLLM and SGLang across 20+ models, reaching over 1,700 tokens/s on DeepSeek V4.1 Flash.

A chart compares prefill throughput of DeepSeek V4.1 Flash at batch size 1 on 2xGB300 TP2 between the zml/llmd dev build and vLLM commit a608b6e6.

TensorFold with EXL3 4.0bpw quant across four Sparks in TP4: 82.8 tok/s decode on code, 2,077 tok/s reading a 32K prompt, 15.8s first token; QA 90.7%, MMLU 93.7%, Tool Eval 93/100.

A user fed six reference images to DeepSeek V4.1 Flash 1017 in Qoder and iterated 62 times, calling its price-performance a beast; compact issues forced new sessions.

ARC-AGI-1 rose from 97.2% to 98.4% with reasoning tokens growing from 26k to 34k per task, while ARC-AGI-2 improved from 75.6% to 79.2% with tokens rising from 108k to 134k.

With unlimited thinking time DeepSeek V4.1 Flash reaches 87% versus GPT-6 Astra at 74% and Muse at 62%, but in real time every large model drops to 0% while a post-trained Qwen 4B hits 90%.

B.AI moves DeepSeek V4.1 Flash to V4.1 Flash-1017 automatically under the same model ID, prompting complaints about missing speed, benchmark, pricing and compatibility notes.

A fully local 12-question test on 2x DGX Spark covering medical stats, paper-reading traps, fake citations and code security review ended 33-32 for DeepSeek V4.1 Flash over GLM-5.3-Flash, essentially a tie.

On a V100 machine with 256GB RAM and 96GB VRAM, the author finds V4.1 Flash ridiculously slow and sticks with Qwen for local runs.

After two weeks of driving DeepSeek V4.1 Flash from Claude for code, the user says backend work is near-flawless and limits drain slower, while UI and visual tasks still go to Opus.