DeepSeek V4.1-Flash-1017 explained: what has changed

DeepSeek has published a new version of its V4.1 Flash AI model, called V4.1-Flash-1017. It is a minor update, but the published scores are higher than the original's, and anyone can download and use the model for free under an open licence.

DeepSeek released the model on Hugging Face, a site where developers share AI models. According to the model card, the repository and the weights (the files that make up the trained model) are under the MIT licence, which is a permissive open-source licence.

The model card describes "substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks", in DeepSeek's words. "Agentic" tasks are ones where an AI carries out multi-step work on its own. DeepSeek also says the model cuts task execution time and output length, which lowers cost.

Here is how Flash-1017 compares with the original Flash on the benchmarks (standard tests) in the card:

The card also compares the model with K3, GLM-5.3, GPT-5.6 Sol and Opus 5. Opus 5 scores higher than Flash-1017 on most of the listed benchmarks.

The context window, meaning how much text the model can consider at once, is 1M tokens. A token is a small chunk of text, roughly a word or part of one.

If you only use chatbots and apps, nothing in the notes says this model is in a product you can use today. What is confirmed is that the weights are free to download, and that DeepSeek's own tests show gains over the earlier Flash model.

Treat the scores with care. They come from DeepSeek's model card, and the same card shows Opus 5 ahead on most of the tests listed.

Some details are not settled. The model card gives no release date. The 17 October date comes from a news roundup on aiweekly.co. That roundup also lists 552B total parameters (a rough measure of model size) and new Search and Fetch APIs priced at $0.75 per 1,000 successful requests. The model page shows 763B, and it lists no API pricing, so treat those roundup figures as unconfirmed.

The model is on Hugging Face, and developers can download it. DeepSeek's recommended settings are:

There is no Jinja chat template, which is the file many tools use to format conversations for a model. Instead, DeepSeek provides a Python reference implementation (encoding.py) and a Rust toolkit called deepseek-recipe for production use. In practice, this is a job for developers rather than casual users.