Deepseek V4.1 Flash 1017 在智能体任务上击败更大对手，同时再次大幅压低单任务成本

Chinese AI company Deepseek has released Deepseek-V4.1-Flash-1017 on its platforms and as an open-weight model on Hugging Face and ModelScope. Trained with reinforcement learning using self-feedback, the model runs faster, produces shorter outputs, and performs well in end-to-end software development and agent tasks, the company says. In internal benchmarks, the update beats larger rivals like Claude Opus 5 and GPT 5.6-Sol on coding and agent benchmarks. Deepseek is also opening up its in-house search engine to developers through new Search and Fetch APIs. Chinese AI company Deepseek has shipped an updated version of its Flash model with major gains in agentic and coding tasks. The company is also opening up its in-house search engine to outside developers for the first time. Deepseek-V4.1-Flash-1017 is available now through the company's website, app, and API. Deepseek also published the model weights as open source on Hugging Face and ModelScope . Compared to its direct predecessor, the model posts better benchmark scores, finishes tasks faster, and produces shorter outputs. The company calls the update a Pareto improvement in the tradeoff between performance and cost. For building complete software projects from scratch, Deepseek uses reinforcement learning with self-feedback (RLSF). This post-training method lets the model use its own internal confidence scores or self-generated critiques as a reward signal. That removes the need for human evaluators or external reward models. Deepseek says the approach improves end-to-end project development for tasks ranging from game programming to 3D modeling and design work. The model can also animate compact explainer videos, similar to the recently released Opus 5.5 . Ad The model was tuned for development environments like Deepseek Harness (version v0.2.1-rc.1 or later recommended), Tencent CodeBuddy, and WorkBuddy. In Deepseek's own benchmarks, V4.1-Flash-1017 outperforms its predecessor V4.1-Flash. It also beats larger competing models like Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3 in several coding and AI agent categories. Ad A cost chart for the DeepSWE benchmark shows how much the shorter outputs matter for efficiency. V4.1-Flash-1017 hits a top score of 77.0 percent at an average cost of just under $0.20 per task. In its low setting, it still clears 72 percent for about $0.10. The predecessor scored 74.2 percent but cost around $0.50 per run. Rival models like GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 land between $2 and over $10 per task in this chart, and none of them match the smaller Deepseek model's accuracy. Independent benchmarks haven't come out yet, and benchmarks only tell half the story anyway. Developer impressions on code quality, for example, can differ wildly from benchmark results. Ad Alongside the language model, Deepseek is making its web search tools (Search & Fetch APIs) available to all developers. The company says these services run on a search engine it built internally that already handles hundreds of millions of queries per day worldwide. Setup details are in the API documentation . Ad Each endpoint costs $0.75 per 1,000 calls. Deepseek API users can access the features for free through server-side tools in Deepseek Harness, Claude Code, and Codex until October 31, 2026 (4:00 PM UTC).