Two announcements this week, both wearing the same word.
On Monday, DeepSeek released DeepInfer — the internal inference stack it uses to serve its models — under the Apache 2.0 license. Real code, really free. And over the weekend into Monday, NVIDIA and the Baker Lab's RFD4-Proteina shipped weights under BSD 3-Clause: a 435.6M-parameter model that designs proteins, antibodies, enzymes, and binders, running on a single H200.
Both are "open" in the legal sense. The licenses are real. The downloads work. If you stop reading there, this week was a victory lap for openness.
Read one step further and the word starts to stretch.
DeepInfer's release is candid about what's missing: no KV cache system, no multi-machine stage separation, no replica load balancing. These are not decorative components. They are the machinery that turns inference code into a serving system — the difference between running a model and running it at the scale and cost that matters. DeepSeek also published pricing for its V4.1-Flash: $0.035 in, $0.12 out per million tokens, against OpenAI's GPT-6 Astra at $10 and $50 — a ~99.65% undercut on input pricing. The release reads, on one level, like a bomb thrown at the margin model of American labs. On another, the withheld parts are exactly where the remaining margin lives.
RFD4-Proteina has a subtler gap. The weights are open, the license permits commercial use, and the model card is unusually honest — the worked examples are explicitly not experimental results, a caveat you almost never see in a launch announcement. But the training-data composition is undisclosed. A flow-matching protein-design model is, in an important sense, the crystallized residue of its training data; not knowing what went in means not fully knowing what came out. Open at the weights layer, closed at the data layer — and the data layer is where the biosecurity questions, the contamination questions, and most of the interesting science live.
Neither of these is a scandal. DeepInfer's authors didn't hide the omissions; Baker's team didn't claim the data was open. But taken together, they show what "open" has become: a license word applied to the artifact, while the gradient — open where, along which axis, and who benefits from what's withheld — carries the real information.
The useful question was never "is it open?" That's binary and legalistic, and it flatters whoever wrote the license file. The useful question is: which parts are open, and are the open parts the parts that would let a stranger actually do the thing?
Take the missing KV cache as a diagnostic. A KV cache is the least glamorous component in an inference stack — a performance optimization, a memory trick. But it's also exactly the component that separates a demo from infrastructure. Withholding it doesn't stop a researcher from studying the architecture; it stops a competitor from matching the serving economics. The withheld parts, in both releases, are the parts that convert capability into deployment. That's not an accusation — it's a pattern worth naming. Open-sourcing the architecture but not the serving layer is giving away the recipe and keeping the oven.
And on the same weekend, the Department of Homeland Security ordered federal civilian agencies to remove DeepSeek products from their systems — a reminder that openness and trust are different axes entirely. Code can be open and still be treated as untrusted. A license says what you're allowed to do with software; it says nothing about whether you should.
The deeper point is about the commons. Openness works as infrastructure only when the open parts include the boring parts — the caching, the load balancing, the data provenance, the evals. A commons that only stocks the showcases isn't a commons; it's a display window. The window is fine — windows persuade people, and persuasion matters — but nobody should confuse it with the warehouse.
So read every "open" release with a second question: what would this release look like if the goal were maximal usefulness to strangers, rather than maximal credit for openness? DeepInfer's answer is genuinely strong — the release notes list the gaps instead of laundering them. RFD4-Proteina's is mixed — honest caveats beside an undisclosed training mix. The honest version of the word is a gradient, and this week two serious teams showed us what the honest version looks like, gaps and all. Grade them on the gradient, not the word.
Loading…