Cactus, an energy-efficient AI inference framework designed for mobile devices, has sparked heated debate in the developer community after switching from the permissive Apache 2.0 license to a restrictive non-commercial license just two weeks ago. The framework, which promises to run AI models efficiently on budget and mid-range phones, now requires paid licensing for commercial use.
The timing and nature of this license change has drawn sharp criticism from developers who view it as a classic bait and switch tactic. Many in the community see this as part of a troubling pattern where startups launch with open-source licenses to gain traction, then pivot to restrictive licensing once they establish a user base.
 |
|---|
| The logo of Cactus, the AI inference framework at the center of the licensing controversy |
Community Trust Concerns
The license change has fundamentally altered how developers view the project's credibility. Users who had integrated Cactus into their applications now face uncertainty about future updates and potential legal implications. One developer expressed particular frustration after building a paid app using Cactus under the assumption it would remain truly open source.
The controversy highlights a broader issue in the startup ecosystem where open source claims are increasingly viewed with skepticism. Community members noted that many projects follow this same pattern: start open source, gain users, then restrict licensing after securing funding or market position.
Technical Performance vs. Business Model Tensions
Despite the licensing controversy, Cactus does deliver impressive technical capabilities. The framework can run a 600MB Qwen3 model at 16-20 tokens per second on older devices like the Pixel 6a and iPhone 11 Pro, with performance reaching 50-70 tokens per second on newer flagship phones. The company has also reported 3x performance improvements in recent updates.
However, the community questions whether these optimizations are truly innovative or simply standard mobile optimization techniques that any experienced developer could implement. This skepticism extends to concerns about the company's long-term strategy and whether the license change is intended to create legal barriers for competitors.
Pricing and Commercial Viability Questions
The framework's commercial pricing remains unclear, with the company stating it uses custom pricing while they calibrate their business model. This lack of transparency has added to community concerns, especially for developers who need predictable costs for their projects. The company describes their fees as tiny but provides no concrete pricing information publicly.
For commercial users, Cactus positions itself as a cost-effective alternative to cloud APIs, claiming savings become apparent after processing approximately two minutes of daily inference. However, without clear pricing, developers cannot properly evaluate the total cost of ownership compared to established alternatives like Ollama or llama.cpp.
Developer Impact and Alternatives
The license change has practical implications for existing users and potential adopters. Developers working on open-source projects under licenses like AGPL3 now face complications when considering deep integration with Cactus. The restriction creates legal complexity that many prefer to avoid entirely.
Open source for the PR, then switching to non-open licensing is a cowardly, bullshit move.
The community has begun recommending alternatives like react-native-ai for developers seeking truly open-source solutions for mobile AI inference. These alternatives may lack some of Cactus's optimization but offer the licensing certainty that developers value for long-term projects.
While Cactus leadership has responded to feedback by clarifying that the license allows free use for personal projects and small developers, the damage to community trust appears significant. The incident serves as a reminder for developers to carefully evaluate the long-term licensing commitments of projects they integrate into their applications.
Reference: cactus-compute/cactus