DeepSeek V4.1 Flash hands-on tests and comparisons

On 133 DACH-Bench tasks on DGX Spark, DeepSeek V4.1 Flash (TP3, 3.5bpw) scored 98.8% quality at 159-164 tok/s with four users, versus 92.7% for Qwen3.8-Flash-Next and 98.3% for GLM-5.3 Flash.

After rewriting its engine in Rust, Tide LM runs native MXFP4 DeepSeek V4.1 on two DGX Sparks at 31.8 tok/s on prose and about 38 tok/s in real use.

On the same 3D Sonic prompt, DeepSeek took 44 minutes for $0.28 while Opus 5.5 took 27 minutes for $5.67, prompting the author to ask whether the 20x pricier Opus is worth it.

A user reports OpenCode Go's DeepSeek V4.1 Flash 1017 loses to the official API every round, is cheaper but slower with frequent timeouts, and looks like a quantized self-hosted deployment.

A user running V4.1 Flash 1017 in DeepSeek Harness says prompting it to use REA with Ghidra let it build nearly a whole app, though the frontend needed the most rework.

A user runs the GSQ-RCO IQ2_XS quant of DeepSeek V4.1 Flash 1017 on an RTX 5090 with R9 9950X and 96GB DDR5 after upgrading Strata from 0.2.4 to 0.2.6, showing the setup without speed numbers.

The author says Kimi 3.1 is a reasonable alternative only when Opus 5.5 or Sol 6.1 are unavailable and the task is too complex for DeepSeek V4.1 Flash, such as 2D/3D design work.

A user complains Grok 4.7 is stale, saying V4.1 Flash 1017 does what Grok does at much lower cost and works well enough in OpenCode Go to justify a Cursor sub.

A user documents copying the deepseek-v4.1-flash-ist-das-2.25bpw pack into Maya with a new config, running across four GPUs with no code or file conversion.

In a same-pack A/B against Project Maya, Strata 0.1.51 on an RTX 5090 ran DeepSeek-V4.1-Flash at 28.7/1013 vs 36.2/1333 tok/s, with the author citing roughly 26% headroom in the prompt kernels.

A team had a coding agent on DeepSeek V4.1 Flash fine-tune a small open model that routes call-centre calls; it shipped with model and traces open.

coralbricks_ai ran a real ML job with Claude Code on DeepSeek V4.1 Flash, fine-tuning a small open model to decide who a call centre should ring, and recorded the whole run.