DeepSeek Launches V4.1-Flash-1017 and Opens Its Search Engine to Developers

Chinese company DeepSeek launched an updated version of its "Flash" model under the name V4.1-Flash-1017, making it available via its website, application, and API, as well as publishing its open weights on Hugging Face and ModelScope.

The company states that the new version is faster than its predecessor and produces shorter outputs, with improvements in end-to-end software development tasks and tasks performed by AI agents. In its internal benchmarks, it outperformed larger models such as Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3 across several coding categories.

In the DeepSWE benchmark, the model scored 77.0% at a cost slightly under $0.20 per task, while its predecessor reached 74.2% at a cost of around $0.50. In a low-cost configuration, the model exceeds 72% at around $0.10, according to figures published by DeepSeek. The company notes that competing models such as GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 range in cost between two dollars and over ten dollars per task. Independent benchmark results for the new model have not yet been released.

DeepSeek relies on reinforcement learning with self-feedback for training, as the model uses its internal confidence scores or self-feedback as a reward signal, reducing the need for human evaluators or external evaluation models. Alongside the model, the company opened its internal search engine to developers via Search and Fetch interfaces at a cost of $0.75 per thousand requests, with free usage available through selected tools until October 31, 2026.

What do these terms mean?

Open Weights: Files representing what the model learned during training; releasing them allows developers to run or modify the model without starting from scratch.

Reinforcement Learning with Self-Feedback: A training method in which the model evaluates its own outputs and uses this evaluation to improve its performance in subsequent attempts.

Application Programming Interface (API): A technical intermediary that allows another program to send a request to the model and receive the result automatically without human intervention.

Task: A unit of work requested from the model to complete, such as fixing a software bug, building a small application, or answering a question.