DeepSeek V4.1 Flash hands-on tests and comparisons

The author reviews this week's Claude token and claude.md threads and argues the usage gap is almost all setup, citing 66M tokens run through DeepSeek V4.1 Flash for $0.55 with Opus orchestrating and cheaper models typing.

Using fraserprice's BRRR recipe with fused decode kernels on b12x vLLM, DeepSeek V4.1 Flash reached 669 tok/s decode at batch 1, 9ms per step, 174ms TTFT and 11K prefill.

llama.cpp build b11870 adds DeepSeek-V4.1-Flash support with Unsloth's Daniel Han as co-author, enabling local offline runs, though required memory is not yet documented.

A video argues DeepSeek mirrors Linux's open-infrastructure playbook: DeepInfer built on vLLM, open 3FS for training data and KV caches, Huawei Ascend tuning, and developers able to run V4.1 Flash themselves.

Max Woolf's analysis finds GPT-5.6 Luna cut 80% and Fable 5.1 cache reads cut 75% between July and September, while DeepSeek V4.1 Flash now has 29 hosting providers and Cloudflare's effective input cost is within 10% of official.

On the same task, V4.1 Flash 1017 cost $1.51 and took 75 minutes while GPT-Sol-6.1 cost $1.78 and took 34 minutes; the author asks readers who won.

A user says paid models exhaust the OpenCode Go subscription in five days, while free DeepSeek V4.1 Flash via @opencrabs ran all their flows for 13 days straight.

Using Goldman's forecast of 120 quadrillion agent tokens per month by 2030, the author calculates that every Grace Blackwell 300 running DeepSeek V4.1 Flash could already produce 13,200 quadrillion tokens a month — 100x demand.

A developer says only Qwen 3.8 Flash Next + Strata is practical on tok/s, while DeepSeek V4.1 is good but too heavy, running under 50 tok/s and needing an RTX PRO 6000 or better.

On 105 real bugs from two overseas projects, Kimi K3.1 Pro scored 21, DeepSeek V4.1 Flash 42.7, and Opus higher, undercutting the idea that pricier models perform better.

All kernels rewritten in TileLang with bit-identical outputs; prefill up 12–14% on four Sparks, 75 tok/s single stream, two logprob bugs fixed, 5,730 tests passing.

On a Python Doom-style game rendered in tmux Braille, DeepSeek V4.1 Flash 1017 at ~40 tok/s on 2×3090 kept producing results while Nanbeige4.3-33B at 70 tok/s on one 5090 overthought.