DeepSeek V4.1 Flash hands-on tests and comparisons

On a roughly $6k Strix Halo plus two RX 7900 XTX setup tuned via Project Maya, DeepSeek V4.1 Flash 1017 reaches 41.3 decode and 1,291 prefill tok/s, with 8.3x faster prefill on a 130K prompt.

First try of deepseek-v4.1-flash-1017 on a local rig hits about 100 tok/s with the Maya-S quant, with a caveat that quantization likely costs two or three benchmark points.

Six community recipes for DeepSeek V4.1 Flash ran on the same four DGX Sparks: all passed 71 of 75 qeval tasks and found every passphrase in a 1M-token prompt, while single-user speed ranged from 77 to 134 tok/s and vLLM served 32 concurrent users at 288 tok/s.

After an analysis blamed his setup and failed requests for TTFT numbers, the author argues the critique ran 148 requests with zero image input, making it benchmaxxed rather than a real multi-agent workflow.

On short prompts, MiniMax-M3.1-Flash led at 0.89s with DeepSeek V4.1 Flash at 0.91s, while GLM, Kimi and GPT-6.1-sol all exceeded 4s.

The author argues 1M context is overkill for most coding tasks and the sweet spot for DeepSeek V4.1 Flash is 300K–500K, recommending a 400K default with manual compaction since agents count max output toward the limit.

The author ran 11 open-weight models locally on a Mac mini; DeepSeek V4.1 Flash and GLM-5.3 Flash looked stable but drifted back toward Lincoln's original wording, suggesting memorization, while the largest models streamed from SSD and took ~36 of 45 hours.

Over 7.3 hours and 1,214 requests on a GB300, the model slept 51% of the time, background evals got 3.7 free GPU-hours, and 85% of requests waited nothing extra (1.3s average added wait).

Using the official API at high thinking and two turns totaling 16M tokens, the author had V4.1 Flash generate a three.js night scene with frogs, a bat and fish, calling the result high quality.

Using DeepSeek V4.1 Flash via Jarvislabs Token Factory inside opencode, the author finds changes appear almost instantly while iteratively building a small game, all recorded on the same VM.

Official DeepSWE v1.1 numbers across five harnesses range from mini-SWE at 74.2% to OpenCode at 65.5%, an 8.7-point spread for the same model.

A comment notes that an engine like Strata can run a DeepSeek V4.1 quant sized for 256GB of memory by offloading portions, leaving some system memory after loading.