DeepSeek V4.1 Flash hands-on tests and comparisons

A Japanese post says V4.1 Flash appears post-trained for Dsh, scores above Astra Fable, ships open weights on Hugging Face, and shows strong 3D CG demos.

Official notice says to call deepseek-flash, now upgraded to V4.1-Flash-1017 with unchanged usage; legacy names like deepseek-v4-flash still work but are served by the 1017 model at Flash pricing.

V4.1-Flash-1017 ships MIT-licensed weights with the same ~550B/8B-16B-active architecture, Terminal-Bench up 31 to 42 and DeepSWE 77 above Opus 5; the checkpoint is about 510GB.

Terminal-Bench 42.3 in the Claude Code harness, DeepSWE 77.0, 1M context and MIT weights, with $0.60/1M output off-peak plus step-by-step Claude Code setup and free Search/Fetch until Nov 1.

User reports that V4.1 Flash 1017, trained with self-feedback RL, improves building complete projects from scratch, especially game development and 3D modeling tasks.

The official release says the new model is tuned for Harness with faster tasks, shorter outputs and lower token costs; users need 0.2.1-rc.1 or later, and first-time users get a 6-yuan credit.

A user recap: gains in game dev and 3D modeling, shorter task time and outputs, a web search API indexing hundreds of billions of pages, and a recommendation to use DeepSeek Harness.

After the official DeepSeek V4.1 Flash release, a user is running their own tests and inviting others to send prompts to try.

Versus the original V4.1 Flash, scores rise in coding, terminal, agent and multimodal tool use, with Terminal-Bench 4.0 jumping from 31.2 to 42.3 and some tasks beating GPT-5.6 Sol.

V4.1 Flash-1017 ships 763B open weights under MIT, self-reported DeepSWE 77.0 vs Opus 5's 74.0, but trails on Terminal-Bench (42.3 vs 51.8) and HLE (42.6 vs 56.3); includes a gummy-bear physics demo.

The post summarizes V4.1 Flash 1017 as open-weight with 1M context, competing with Opus 5 on coding and agent tasks at a fraction of the cost.

Official benchmarks show V4.1-Flash-1017 keeps its pricing while matching GPT-5.6 Sol overall and surpassing it on some coding and agent tasks.