Deepseek Releases V4.1 Flash 1017 Open Model
Deepseek has launched its open-weight Deepseek-V4.1-Flash-1017 model, outperforming larger rivals on coding benchmarks while drastically lowering per-task execution costs.
Chinese AI firm Deepseek has launched Deepseek-V4.1-Flash-1017 across its web platforms, mobile app, and API, while publishing model weights on Hugging Face and ModelScope. Trained using reinforcement learning with self-feedback (RLSF), the system uses internal confidence scores and self-generated critiques as rewards instead of human evaluators or external reward models. This approach enables the model to complete end-to-end software projects, including game programming, 3D modeling, design work, and short explainer video animation.
In Deepseek's internal tests, V4.1-Flash-1017 outperformed its predecessor as well as larger proprietary systems like Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3 in agentic and coding evaluations. On the DeepSWE benchmark, the new release achieved a peak accuracy score of 77.0 percent at an average cost below $0.20 per task, while a low-cost configuration reached 72 percent for roughly $0.10. By comparison, the prior V4.1-Flash scored 74.2 percent at $0.50 per run, while competing systems such as GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 cost between $2 and over $10 per task.
Alongside the model update, Deepseek launched its Search and Fetch APIs to give external developers access to its proprietary web search infrastructure, which handles hundreds of millions of daily queries. The search endpoints cost $0.75 per 1,000 calls, though developers can access them at no charge via server-side tools in Deepseek Harness, Claude Code, and Codex through October 31, 2026, at 4:00 PM UTC.
For software engineers and AI practitioners, the update offers a dramatic cost reduction for automated software generation and agent workflows. Deepseek described the update as a Pareto improvement in performance and cost trade-offs. The model is specifically tuned for environments including Deepseek Harness version v0.2.1-rc.1 or later, Tencent CodeBuddy, and WorkBuddy. However, developers should note that potential Chinese regulatory changes might tighten restrictions on publishing open-weight models that rank among the top ten global systems.
This is our own summary of reporting by The Decoder
