Run DeepSeek V4.1 Flash (1017) in the EU through an OpenAI-compatible API

DeepSeek V4.1 Flash was released in September 2026, followed on 17 October by DeepSeek-V4.1-Flash-1017, a minor update with better reasoning, coding, agent and visual results and shorter outputs. We treat both as one model and serve the 1017 version. It has 552 billion backbone parameters and activates 8 billion parameters per token during prefill and 16 billion during decoding.

It uses a new causal encoder-decoder design and a second generation of compressed sparse attention, which DeepSeek says cuts the cache needed per token to about a quarter of V4 Flash. A vision encoder trained from scratch lets it read images natively. The context window is one million tokens, and the weights are MIT-licensed.

V4.1 Flash is aimed at agent work: long coding sessions, multi-step tool use and tasks that mix screenshots, charts and text. Its reasoning effort is a number from 1 to 100, so you can tune cost against accuracy finely instead of picking from three levels. Long inputs are cheap to process because so few parameters are active during prefill.

At the top of the effort scale it produces long reasoning traces, so output cost and time rise quickly. It produces text, not images. Like all DeepSeek models, its answers reflect its maker's training choices, including on politically sensitive subjects. For pure text tasks at the lowest cost, DeepSeek V4 Flash or a smaller model may be enough.

inference.eu uses the OpenAI request format, so the OpenAI SDKs and most agent tools work unchanged. Set the base URL to the inference.eu API, use your inference.eu key and set the model to deepseek-v4.1-flash.

Send images as image_url content parts, interleaved with text if you like. Set reasoning_effort to an integer between 1 and 100; low values answer quickly, 100 reasons as far as the model goes. DeepSeek recommends temperature 1.0 and top_p 0.95. DeepSeek ships its own prompt encoding rather than a chat template; we handle it for you. The Anthropic Messages format is also available.

DeepSeek's own app and API process data in China, which is a hurdle for most European companies. With the open weights that hurdle disappears. On inference.eu, DeepSeek V4.1 Flash runs on infrastructure in the European Union: prompts, images and responses are processed in EU data centres, are not sent to DeepSeek, and are never used to train any model.

Every customer gets a data processing agreement under Article 28 of the GDPR and our list of subprocessors, so you can show that inference involves no transfer to a third country. Billing is in euros. The MIT licence allows commercial use with minimal conditions.