DeepSeek V4.1 Flash hands-on tests and comparisons

Official estimates for a $10 OpenCode Go plan per 5 hours: Kimi K3.1 Pro ~147 requests, DeepSeek V4.1 Flash ~34,670, MiniMax M3 ~3,200.

The author releases four abliterated builds quantized with a custom calibrated lecore VQ 11-bit format, improving perplexity from 3.62 to ~2.8 and top-1 agreement from 74.5% to ~81%, runnable only on a custom ds4 runtime.

The poster claims that on subscription pricing, Kimi K3.1 Pro costs more than 20 times DeepSeek V4.1 Flash 1017, and asks users to confirm.

On 105 planted bugs across two production repos, DeepSeek V4.1 Flash 1017 fixed 42.7, slightly above Opus 5.5's 41.7 and Kimi K3.1 Pro's 36, at roughly $1–2 cost.

The author is running a DeepSeek V4.1 Flash 1017 quant on Strix Halo 395+ at about 3.90 effective bpw, favoring quality with acceptable prefill, and targets over 4.00 bpw next.

Strata 0.2.4 runs DeepSeek V4.1-Flash on one 12GB GPU with 46–64GB RAM, keeping hot experts in memory and streaming the rest from SSD.

Qoder's personal plan offers DeepSeek-Flash (V4.1-Flash-1017) at 0x cost, free through Nov 1 midnight, with max thinking intensity and no credit deduction.

A write-up summarizes NVIDIA's DGX Spark ring-vs-switch TP4/TP8 throughput numbers on Qwen3.8-27B and adds the author's earlier DeepSeek V4.1 Flash TP4 ring results, where relaying RoCE cut decode step time and lifted concurrent output 13–18%.

A user runs unmodified V4.1 Flash weights on four 2019-era GPUs, hitting 314 tok/s peak decode on real coding work and ~200 tok/s average in coding agents.

A user is testing TensorFold patches that let sm86 and sm121 machines cooperate over RoCE, pooling RAM and VRAM to run DeepSeek V4.1 Flash.

Same prompt, harness, and reasoning effort compared across two providers: DeepSeek Platform ran 1h30m at 213 tok/s with 99.8% cache hits, OpenCode Go 1h18m at 143 tok/s with 98%.

Having run qwen3.8-flash-next locally on strata without much use, the user wants someone to get DeepSeek V4.1 Flash 1017 running locally to minimize AI spending.