DeepSeek released DeepSeek-V4.1-Flash-1017 five weeks after introducing the base DeepSeek-V4.1-Flash on Sept. 10. According to the company, the new version focuses on end-to-end project development “from scratch,” including game creation and 3D modeling, and is optimized for DeepSeek Harness.

The company positioned the release as offering flagship-level performance in agentic tasks at Flash-tier prices. Users describe the intended result, and the model handles tasks ranging from setting up the project structure to producing a finished application.

Benchmark results

DeepSeek compared the release with its previous model and Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5 across 11 tests. All results came from the company’s internal testing. Developers used the Claude Code shell for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE v1.1.

The DeepSWE v1.1 score increased to 77.0 from 74.2, while the Terminal-Bench 4.0 result rose to 42.3 from 31.2. Gains were smaller in visual tool-based tests: Chartography increased to 83.0 from 78.9, BabyVision to 92.5 from 89.6 and ZeroBench to 51.0 from 49.0.

DeepSeek-V4.1-Flash-1017 outscored Claude Opus 5 in DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench. Anthropic’s model led in the other seven tests, with the largest gap — nearly 14 points — in Humanity’s Last Exam. DeepSeek also published an accuracy-to-cost comparison for DeepSWE to support its claim of a Pareto improvement.

Availability and pricing

The model now runs by default and without charge in DeepSeek’s web chat and mobile app. API users continue to access it through the deepseek-flash identifier. The legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp names remain functional, but requests are served by version 1017 on the Flash pricing tier.

API rates remain unchanged. Per 1 million tokens, off-peak prices are $0.003 for cached input, $0.15 for uncached input and $0.60 for output. Peak rates are $0.006, $0.30 and $1.20, respectively. Peak periods run from 01:00 to 04:00 UTC and 06:00 to 10:00 UTC on weekdays; weekends and Chinese public holidays are entirely off-peak.

The weights are available on Hugging Face and ModelScope under the MIT license. However, the source said a model with hundreds of billions of parameters requires server-grade hardware rather than a typical home computer.

Test in DeepSeek Harness

Incrypted tested the model by asking it to build a self-contained Three.js index.html file containing an interactive 3D model of Beijing’s Forbidden City. The prompt required major buildings along the central north-south axis, orbit controls, a guided camera tour, day and night modes, lantern lighting and labels shown on hover.

The test used DeepSeek Harness, DeepSeek’s open agent shell for Windows and macOS. The recommended builds were v0.2.1-rc.1 and v0.2.1-rc.2, available from the project’s GitHub releases page. Users can log into a DeepSeek account or connect a key from the DeepSeek platform, select DeepSeek-V41-Flash and retain the default High reasoning level.

The completed file included the moat, gates, courtyards, three principal halls on marble terraces and four corner towers. Its guided tour moved the camera south to north, while the day control switched the scene to night lighting. Harness showed the model launching four sub-agents to check syntax and JavaScript behavior. The final session recorded three runs, 96 steps, 26.3 million tokens and a 99% cache-hit rate.

Limitations and restrictions

DeepSeek did not disclose exact reductions in task completion time or response length, and independent measurements were limited at publication. The Harness builds recommended by the company are release candidates and may contain bugs or experimental features. In Incrypted’s Windows test, PowerShell did not launch, forcing the agent to work without a terminal.

On Oct. 19, the U.S. Department of Homeland Security required federal civilian agencies to remove DeepSeek products from their systems and barred the company from participating in government contracts. The restriction does not apply to private users or businesses, though companies contracting with the U.S. government should consider it when selecting a model.

Source: Incrypted