DeepSeek V4.1 Flash hands-on tests and comparisons

In a simulated city nuclear-strike task, DeepSeek V4.1 Flash 1017 cost $0.67 and took 2h27m, far below Opus 5.5 at $148.65 and nearly 12 hours.

A Portuguese reply argues the 5090 price was already charity from Nvidia, and that a GB300 system recoups its cost in roughly five days serving DeepSeek V4.1.

Infra providers are racing to speed up V4.1 Flash 1017, from DeepSeek's official 333 tok/s to Fireworks AI's 1,521 tok/s, enabled by open weights and DSpark speculative decoding.

One prompt to DeepSeek V4.1 Flash-1017 produced a self-contained HTML file where the phone, board, legal moves, camera cuts and timing are all hand-coded SVG and CSS animation, with no assets or tool calls.

Project Maya v1.1.22 runs DeepSeek-V4.1-Flash-1017 fully in 128GB; one dev streams experts off SSD on a single RTX 3090 plus 64GB RAM at about 4 tok/s.

On an M3 Ultra with DeepSeek Harness, V4.1 Flash-1017 hit 16/16 in one run at reasoning 40 with a 9k budget, but output consistency dropped and 3 of 6 runs hit the 210-minute timeout.

Same four-stage brief for a three.js dog: V4.1 Flash cost $0.22 in 29 minutes with a leg clipping bug, while Opus cost $2.43 in 67 minutes with a one-eyed dog.

On the Vals Index the tested model costs $0.97 per test, while GPT-6 Luna ($0.45), DeepSeek V4.1 Flash ($0.66) and MiMo V2.6 Flash ($0.22) are all cheaper and score higher.

The same real agent sessions ran 24 hours across OpenRouter, Vercel, Ramp and Keln, measuring cache hit rate, speed, reasoning and cost; Keln led cache hits at 94% while reasoning numbers were worse.

Project Maya splits the 510GB model across GPU core layers and hot experts, system RAM and on-demand SSD reads, reaching about 4 tokens per second on an RTX 3090 with 64GB RAM.

Local deployment report: three V100 32GB cards plus DDR4 hit ~70 tok/s decode and ~446 tok/s prefill at 256k context using the new Maya-XS quant.

Testing V4.1 Flash on DwarfStar gave 18 tok/s, slower than llama.cpp's 27–31 tok/s; after tuning it reached 27 tok/s, matching the vcruz build.