Intel / article

Latest open artifacts (#25): DeepSeek V4.1 Flash 1017, American-made models, and new regulatory controls in the US & China

Source: interconnects.ai
Author: Florian Brand
Date: Published Oct 20, 2026

Context

This issue of Interconnects' open-model roundup centers on DeepSeek because a new model, a new restriction and a new open-source release arrived close together. The model, DeepSeek-V4.1-Flash-1017, is one of two new iterations at a new size, listed as 552B-A16B. That notation usually means 552 billion total parameters with about 16 billion active for each token. The model accepts image inputs and uses the MIT license. The authors call the license choice notable at a time when model licenses are becoming more restrictive. They say the model has drawn for its performance at a low price and show it leading a price-versus-performance chart on DeepSWE, an agentic coding benchmark.

On policy, the post says a Department of Homeland Security order bars DeepSeek and its models from being vendors in federal civilian procurement. The authors say it is unclear whether the order reaches fine-tuned models or models trained on DeepSeek outputs, and they point to NVIDIA's Nemotron recipes, which have relied heavily on DeepSeek models. China's Ministry of Commerce also proposed requiring government pre-approval before frontier AI models are shared internationally. The authors consider the reaction overblown because the document says open-source technologies will get automatic approval, though others see a framework for future control.

DeepSeek also published DeepInfer, the software it uses to serve its own models. The authors expect providers to adopt it, which would lower prices for DeepSeek models or raise providers' margins. The post presents the claim that DeepSeek makes up to 90% gross margin on this stack as an outside estimate, not a company figure. The repository says the release covers offline inference and leaves out pieces a complete service needs, including a system and load balancing. This explanation covers the free portion of the post; the rest is for paid subscribers.

Terms in this piece · Glossary

attention — The mechanism that lets a model weigh which earlier words matter for the word it's currently processing — the core operation of a transformer.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
KV cache — The memory a model keeps about text it has already read, so generating each new token doesn't require reprocessing the whole conversation.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.

Read the source www.interconnects.ai