Open models, and where they come from
In one week Reflection released a strong open model, DeepSeek updated its own, and Washington banned DeepSeek from its agencies. Quality is no longer the question. Origin is.
For a company that wants an AI of its own, the most important news of the past week was not a new chatbot. It was three things that happened within a few days: an open model from Reflection AI that fits on a single server, new weights for DeepSeek’s open model, and an order from the US government that removes DeepSeek from its civilian agencies. Together they say something simple. The open models are now good enough for serious work, and the question has moved from quality to origin.
Beam: fewer tokens, not cheaper ones
On 16 October Reflection AI released Beam, its first open-weight model, under the Apache 2.0 licence, which allows commercial use and changes. It has 501 billion parameters, of which 23 billion are active for each token. In 8-bit or 4-bit form it fits on one server with eight GPUs, such as H100, H200 or GB300. The full version needs “tens of GPUs”, VentureBeat reported.
Reflection’s pitch is not cheaper tokens but fewer of them. During training it penalised the model for every token it did not need, and in VentureBeat’s reading of its chart Beam uses about 25'000 tokens per attempt on a coding test where GLM-5.2 needs about 80'000. It is not the strongest open model. On Terminal-Bench 2.1, VentureBeat reported about 80 percent for Beam, against 90.6 percent for DeepSeek V4.1 Flash and 85 percent for Kimi K3. Reflection says Beam will “win on merit rather than on anything else.”
DeepSeek: stronger, and MIT licensed
In the same days DeepSeek published updated weights for V4.1 Flash, dated 17 October in their name. The model card calls it “a minor update”, with gains in reasoning, coding, agentic tasks and vision, and recommends a context of one million tokens. The licence could not be more open: “This repository and the model weights are licensed under the MIT License.”
DeepSeek has also been consolidating. In September it moved all requests for its larger V4 Pro over to V4.1 Flash, “at V4.1-Flash rates”, and began to phase V4 Pro out. By the benchmark VentureBeat cited, it is the strongest of the open models named here.
Washington draws a line
On 19 October CyberScoop reported that the US Department of Homeland Security is barring civilian federal agencies from using DeepSeek products, under a supply chain law from 2018. Agencies must remove the versions they have and may not buy new ones. The reasons are about data: DeepSeek stores data on servers in China, and its prompt and response process “may result in the transfer of sensitive data” to the Chinese government. The order names only DeepSeek, although other Chinese open models such as Kimi and Qwen have drawn similar concerns.
The report does not say whether the ban also covers DeepSeek’s weights when an agency runs them on its own servers, where no data travels anywhere. That open question is the one every company should note. Running a model at home answers where the data goes. It does not answer where the model comes from.
What it means for a model of your own
Open weights on your own servers solve the most obvious problems: the data stays home, and nobody can change or withdraw the model on a Tuesday. They do not solve all of them. A model is also a product with an origin, a licence, training data and behaviour that nobody outside the lab can fully inspect. For a bank, an authority or a defence supplier, the question is no longer whether an open model is good enough. It is whether its origin will pass the next audit, the next customer questionnaire, or the next order like this one.
In practice that means three things. Choose the base model as deliberately as you choose the licence: Apache 2.0 and MIT are both permissive, but neither says anything about provenance. Keep the model replaceable, so that the company’s knowledge, rules and tests live outside it and a better or safer model can take its place in weeks. And test every candidate on your own work before it goes live, not only on public benchmarks.
How Eigna handles it
This is how Eigna is built, the private AI that we set up for clients as a partner of Eigna Labs. It starts from an open model, adapted to the client’s domain and trained on its own work, rules and language, running in its own cloud or on its own servers. A new version goes live only when it tests better than the last one, and it can be switched back in one step. Which open model sits underneath is a decision we make with each client, origin included, and one that can change when a better one arrives.
