DeepSeek V4.1 Flash hands-on tests and comparisons

The poster links OpenRouter's pricing and benchmark page for V4.1 Flash 1017 and complains the listed speed is far slower than their actual experience.

The author published a 3-bit quantized pack for V4.1-Flash-1017 that runs locally at about 100 tok/s, their first Hugging Face repo.

A Japanese post rounds up DeepSeek V4.1-Flash-1017's release facts: 77.0% on DeepSWE v1.1, about $0.19 per task, MIT-licensed weights released same day, and 1M-token context support.

A news-style relay of official claims: improvements in speed, efficiency, reasoning, coding and agent ability, framed as a challenge to OpenAI, Google and Anthropic.

Says the model is out on Hugging Face with 'wish-driven programming' as its pitch, sub-second code generation, speculative self-repair loops and flash pricing.

A Japanese roundup reports DeepSeek shipped V4.1-Flash-1017 across web, app and API on Oct 17, lifting its terminal-task score from 31.2 to 42.3 with shorter outputs and unchanged API pricing.

The post lists V4.1-Flash-1017's MIT license, Terminal-Bench 4.0 gain from 31.2 to 42.3, shorter outputs, unchanged price and 1M context, asking whether open models are closing the coding-agent gap.

A user notes official throughput of roughly 200-300 tok/s and wonders how good the official and Ollama Fetch/Search APIs are, an early observation rather than a benchmark.

A roundup says V4.1 Flash was silently swapped to the 1017 build across web, app and API at unchanged pricing, with 763B parameters under MIT license, 1M context, and Search/Fetch APIs at ¥5 per 1,000 calls, free inside Claude Code and Codex until Nov 1.

Citing official figures, the post notes unchanged API pricing with Terminal-Bench 4.0 up from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0, plus shorter outputs and faster task completion.

From a single three.js prompt, V4.1 Flash 1017 produced an FPS with 5 survival waves, 3 agents and 3 difficulty modes in 37 minutes for about $0.10.

A user asked V4.1 Flash 1017 to render a realistic stormfront and praised the quality of the output.