DeepSeek V4.1 Flash hands-on tests and comparisons

A single M5 Ultra 256GB runs 3-bit quantized DeepSeek-V4.1-Flash-1017 in mlx-serve at 102–112 tok/s (peak 136.7), 1770 tok/s prefill, ~9s to first token at 16k context, usable from Claude Code and Codex.

With identical 2.9 bpw weights and clocks capped at 2,000 MHz, 3 Sparks (TP3) vs 2 (TP2) gave 117 vs 88 tok/s on code, 82 vs 61 on prose and 156 vs 120 on 4 streams, roughly 29-34% faster.

On the same keyboard-building prompt, Opus 5.5 delivered labeled glossy keycaps, ripples and a WPM counter for $0.20, while DeepSeek V4.1 Flash cost $1.70, produced flat boxes and mistyped "hello jelly opus" as "hllo jely opus".

The author wires Hermes Agent to WorkBuddy's free DeepSeek V4.1 Flash through a local magpie gateway, runs multi-hour jobs for two days at zero credits, and shares rate-limit and quota-check lessons.

A user says DeepSeek V4.1 1017 Flash replaced V4 as their main model, with GLM 5.3 Flash as backup, each running well on two DGX Sparks or a Mac Studio.

The new quant upgrades the highest-error gate/up layers to 4-bit, improving KL vs DeepSeek FP8 by 4–20% over 3.5bpw, and is on Hugging Face.

The author reruns community-built recipes on his own hardware, including Mia's V4.1 Flash setup, a four-Spark ring, and Project Maya on a single RTX 5090, posting numbers then results.

Across 89 runs of identical DeepSeek V4.1 Flash 1017 weights in 9 harnesses, scores ranged from 31.8 (Kilo Code) to 40.4 (pi, grok build cli), with Claude Code at 36.4 — the gap comes from the wrapper.

Third-party platform engy sets new DeepSeek V4.1 Flash pricing at $0.03 input / $0.12 output, cutting input 25% and raising output 50%, still 80% below official rates.

Fireworks AI's Ultrafast platform reports DeepSeek V4.1 Flash at over 1521 tok/s, roughly 7x faster than Opus 5.5, pitched at low-latency coding.

The author improved Nvidia Flare for DeepSeek V4.1 Flash tp=3 deployments on DGX Spark, adding three-Spark support and disk caching.

A code read of DeepInfer: Python on a C++ core with TileLang kernels, vLLM-derived, serving V4.1 Flash with 64 GPUs for prefill and 64 for decode, experts over NVLink and FP4/FP8 compressed KV cache.