DeepSeek V4.1 Flash hands-on tests and comparisons

A developer ran the same rocket-launch task on DeepSeek V4.1 Flash 1017 and GPT-6.1 Sol, reporting roughly 8x lower cost and two minutes faster, alongside official MIT-license and DeepSWE 77.0% figures.

Same video task: V4.1 Flash cost $0.78 and took 3 minutes, while GPT-6.1 Sol cost $6.10 and took 5 minutes; the author notes DeepSeek added camera shake.

A Japanese user notes V4.1 Flash is faster, cheaper and stronger with 3.9M views on the launch post, then speculates pessimistically about the US-China AI race and Anthropic's distillation claims.

TNG's local B300 deployment with the system prompt disabled has the model identifying itself as Anthropic's Claude, while other posters saw it claim OpenAI.

A user reports workbuddy started limiting V4.1 Flash 1017 and OpenCode became less usable, asking where to find unlimited free DeepSeek quota.

Official release details: MIT-licensed weights, Harness 0.2.1, and Search/Fetch API at $0.75 per 1K calls; DeepSeek claims up to 77.0% on DeepSWE at under $0.20 per task, with no independent reproduction yet.

A user asked DeepSeek V4.1 Flash 1017 in its chat platform to build a travel website; it took three requests to get everything working, at zero cost, with a video of the result.

Only 37 days after the previous version, the 552B-parameter model is released under MIT on Hugging Face for commercial use, scoring 77.0 on DeepSWE v1.1 versus Claude Opus 5's 74.0.

Local deployment test shows the EXL3 quantized V4.1 Flash running stably around 50 tok/s on two DGX Spark units, with prompt cache working.

Comparison finds official-channel first-token latency around 1.5s versus Cline Pass starting at 2.5s and reaching 47s at 335K context, while output speed stays near 265 tok/s.

Using lucebox RAM pinning, expert tuning and DSpark Q2 quantization, the model reaches up to 14 tok/s on an R9700 Pro, 9070XT, 64GB RAM and 7700X, with only a 32k context window so far.

The author prompts V4.1 Flash 1017 in one sentence to produce a video explaining how an elevator works, with voiceover and background music, benchmarked against Opus 5.5.