The Pentagon learned what vendor dependency costs
In March 2026, the Pentagon banned Anthropic from Project Maven and gave Palantir 180 days to rip it out of core defense systems. The reasons are classified. The fact remains: when your AI depends on a vendor's willingness to serve you, someone outside your chain of command can pull the plug.
Any proprietary AI model accessed via API comes with terms of service the vendor controls. Those terms can change. Export controls can shift. A vendor can exit the defense market, get acquired, or face pressure to restrict military use. The model weights sit on someone else's servers, governed by someone else's legal team. The DoD AI Strategy memo directs AI that works 'on-board, in real time, and often without connectivity.' That directive is incompatible with models you do not own.
Apache 2.0 changes the ownership equation
The Apache 2.0 license grants a perpetual, worldwide, non-exclusive, royalty-free license to use, reproduce, modify, and distribute. Once you download the weights, they are yours. No EULA revocation. No usage restrictions. No export control surprises after deployment.
The most capable Apache 2.0 models as of June 2026: Google's Gemma family, Alibaba's Qwen, and Mistral. Gemma 4 12B is encoder-free multimodal with native audio, runs quantized in 8GB, Apache 2.0. A defense organization that downloads it today has permanent, unrestricted access to a multimodal model. No contract renewal. No API key. No vendor relationship required after download.
What runs on government hardware is what matters
A model on a government-owned MacBook, stored on a government-owned SSD, with weights downloaded under an irrevocable license, works regardless of what happens to the vendor, the internet, or the relationship between the US and the model provider's country of origin.
A naval vessel in the Persian Gulf cannot call an API. An SOF team in a denied environment cannot authenticate to a cloud service. An airgapped SCIF cannot reach any external endpoint. In all three cases, the only AI that works is AI physically present on the device. Open-weight models under permissive licenses are the only models that satisfy all three requirements simultaneously: local execution, no connectivity required, no revocation possible.
The capability gap closed faster than procurement can track
The Edge Mobile LLM Leaderboard tracks models optimized for on-device inference. Models that required 32GB in 2024 run quantized in 8GB in 2026. Models that needed a data center GPU in 2025 run on laptop Apple Silicon. Gemma 4 12B runs multimodal inference on a 16GB laptop.
Defense procurement cycles run 18-36 months. The open-weight ecosystem moves in weeks. By the time a program office evaluates, selects, contracts for, and deploys a proprietary model, multiple generations of open-weight alternatives will have shipped under Apache 2.0 with comparable performance. EdgeLance's model governance architecture treats model selection as a mission planning decision, not a procurement decision. When a better model ships, it goes into the next mission pack.
Open weight does not mean ungoverned
The objection is usually governance: who approved this model, what was it trained on, has it been evaluated for bias and safety. Legitimate questions. They apply equally to proprietary models, where the answers are often less transparent because the vendor controls the information.
EdgeLance applies the same governance chain to open-weight models that any responsible program applies to software components. Models are evaluated before mission approval. Approved versions are pinned and signed. Deployment happens through fleet management with full audit trails. The difference: with open-weight models, the government has full visibility into architecture, training methodology, and weight values. With proprietary models, governance depends on whatever the vendor chooses to disclose.
The Maven/Anthropic episode is a warning. As AI becomes more central to defense operations, vendor dependency risk grows. Open-weight models under Apache 2.0 eliminate that risk entirely. EdgeLance was built on this principle from day one. Every model in the mission loadout is open-weight, locally executable, and irrevocably licensed.