DeepSeek V4.1 Flash hands-on tests and comparisons

A 24-hour AI news digest noting V4.1-Flash-1017 ships MIT weights, activates up to 16B params at output, costs about $0.2 per task, and that its 77.0 DeepSWE score is self-reported and unreproduced.

A Japanese post covers the Oct 17 V4.1-Flash-1017 release, citing 77.0 on software repair and 60.5 on business automation, beating GPT and Claude on some benchmarks with unchanged API pricing, while noting the figures are vendor-reported.

A user says V4.1 Flash outputs at 200 tok/s with good results, yet end-to-end task time matches or trails GPT-6.1 Sol at 40 tok/s, suspecting overthinking.

A side-by-side video has both models build an interactive 3D jet engine across six rounds with the same camera and moves: cutaway, airflow, throttle, signature move and guided tour.

Users tested DeepSeek V4.1 Flash 1017 building MOC bricks in LDraw Nova and posted screenshots of the results for others to judge.

After a morning with V4.1 Flash 1017, the author says it feels like using 6.1 Sol and is strong enough to doubt it came from DeepSeek.

The author calls it the top open-source model, with coding clearly improved beyond GPT-5.6 Sol, and says the previous V4.1 Flash already handled daily company and reverse-engineering work well.

With a portfolio-style prompt, the model produced code for a 20-second motion graphics short in 26 minutes, which the user calls strong and cheap for motion design.

On the same task Opus 5.5 finished in 20 minutes for $6, while DeepSeek V4.1 Flash took an hour but cost only $0.23 — roughly 26x cheaper yet slower.

The new model is fully rolled out to web, app and API with faster inference and continued price cuts, framed as a China-vs-US price war.

The author used local Qwen plus the GLM5.3 API to find OSS vulnerabilities in Docker, and warns that models like V4.1 Flash 1017 can be weaponized for attacks.

Roundup of official details: higher benchmarks and shorter outputs, RL-trained full-project generation, Search & Fetch APIs at $0.75 per 1,000 calls, free in Harness/Claude Code/Codex until Oct 31.