DeepSeek V4.1 Flash hands-on tests and comparisons

A user had V4.1 Flash generate a 3D FPS game entirely in code via DeepSeek Harness, with no external assets, nearing Opus 5.5 and beating other open models.

The author tested V4.1 Flash-1017 on motion design, ranking it just below Claude models and above GPT-6, with a companion motion skill due next week.

In a non-recursive peacock test, V4.1 Flash 1017 lands roughly at GPT-6 Luna Max level, though legs bend oddly, tail feathers look like eyes on sticks and the bike goes nowhere; the run cost $0.53.

A user reports the new release tops Opus 5 on several benchmarks, hits a new cost-performance Pareto frontier across harnesses, and improves game and 3D project building.

Same prompt at max reasoning: Argon took 13 minutes and $1.2, DeepSeek v4.1 took 50 minutes and $0.4, with the author calling the outputs close and DeepSeek's water detail slightly better.

In a same-task comparison with similar runtime (mostly rendering), DeepSeek's full test cost just 0.44 cents, and the author credits free web search and native multimodality for its edge.

Roundup of the release: 763B open-weight model on Hugging Face under MIT, 1M-token context, DeepSWE v1.1 up from 74.2 to 77.0, with shorter outputs and faster agent runs.

WorkBuddy Intl has switched to V4.1-Flash-1017 with free access until Oct 31; DeepSeek says the release is faster with shorter output and tuned for Harness, CodeBuddy and WorkBuddy.

A user reports V4.1 Flash produced a 3D meadow with swaying grass, trees, dandelions and ambient audio from a single prompt, noting it is only a Flash model.

Third-party platform RunInfra offers day-zero access with 327 tok/s output, 30ms to first reasoning token, 1M context, cache pricing, and over 98% cache hit rate on agentic workloads.

The author says the existing vLLM recipe should carry over to the 1017 release by swapping the model, links a DGX Spark repo, and is working on a TensorFold recipe.

Comparing old and new V4.1 Flash in the same DSH setup with an official key, the author's white-vs-green screenshots show a clear improvement in frontend output.