Skip to main content

0004. Dual MIT/Apache-2.0 Licensing with Separate Model License

Date: 2025-11-17 Status: Accepted Deciders: AGX Core Team Tags: licensing, legal, open-source, ai-models

Context​

AGEniX consists of multiple components with different licensing considerations:

Software Components​

  • Core infrastructure (AGX, AGQ, AGW): Rust binaries, libraries, CLI tools
  • Agentic Units (agx-ocr, future AUs): Standalone tools with AU contracts
  • Central documentation (agenix repo): Architecture, specs, contracts

AI Models (Current and Future)​

  • Planning models (Echo, Delta): Fine-tuned models for AGX planning
  • AU-embedded models (e.g., DeepSeek-OCR in agx-ocr): Models bundled with AUs
  • Future distributed models: Models downloaded/cached by users

Requirements​

  1. Maximize adoption: Permissive license to encourage ecosystem growth
  2. Rust ecosystem compatibility: Align with Rust community standards
  3. Patent protection: Protect contributors and users from patent trolls
  4. GPLv2 compatibility: Allow use in GPLv2 projects (if needed)
  5. Commercial-friendly: Enable commercial use without barriers
  6. Model distribution clarity: Separate licensing for AI model weights
  7. Attribution requirements: Ensure credit to contributors

Constraints​

  • Cannot use copyleft (GPL) - prevents commercial adoption
  • Must be OSI-approved for core software
  • Model licenses may have use-based restrictions (acceptable)
  • Must be simple for contributors to understand

Decision​

We will use dual MIT/Apache-2.0 licensing for all AGEniX software code, with separate model licenses for AI weights.

Software License: MIT OR Apache-2.0​

All repositories containing code (Rust, scripts, tooling) will be dual-licensed:

Licensed under either of:
 * Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
 * MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Users can choose which license to apply.

Model License: Custom Per-Model​

AI model weights will have their own licenses:

  1. Upstream model licenses (e.g., DeepSeek Model License)

    • Preserved for models we use/redistribute
    • Documented in MODEL-LICENSE file alongside weights

  2. AGEniX-trained models (future)

    • Custom license similar to DeepSeek: permissive with use-based restrictions
    • Separate from software license
    • Documented in MODEL-LICENSE in model repositories

Documentation License: CC-BY-4.0 (Optional)​

For pure documentation (guides, tutorials), we may use Creative Commons Attribution 4.0 International (CC-BY-4.0) to allow broader redistribution.

Core architecture docs remain under software license (MIT/Apache-2.0) as they're part of the product.

Alternatives Considered​

Option 1: MIT Only​

Pros:

  • Simplest license (very short, easy to understand)
  • Most permissive
  • GPLv2 compatible
  • Highest adoption on GitHub

Cons:

  • No explicit patent grant
  • No protection from patent trolls
  • Less protective for contributors

Option 2: Apache-2.0 Only​

Pros:

  • Explicit patent grant (protects users and contributors)
  • Well-understood in enterprise
  • Contributor license agreement built-in
  • Strong in Rust ecosystem

Cons:

  • GPLv2 incompatible (rare issue, but exists)
  • Slightly more complex than MIT
  • Longer license text

Option 3: Dual MIT/Apache-2.0 (CHOSEN)​

Pros:

  • Best of both worlds: Users choose based on their needs
  • Rust ecosystem standard: Aligns with Rust project itself
  • Maximum compatibility: MIT for GPLv2 projects, Apache-2.0 for patent protection
  • Patent protection available: Users opting for Apache-2.0 get patent grant
  • Simple for contributors: Standard Rust contribution model

Cons:

  • Two license files to maintain (minimal overhead)
  • Slightly more complex explanation for new contributors

Option 4: GPL/AGPL (Copyleft)​

Pros:

  • Forces derivatives to stay open-source
  • Prevents proprietary forks

Cons:

  • Prevents commercial adoption (major barrier)
  • Not compatible with Rust ecosystem norms
  • Restricts downstream use (against our goal of ecosystem growth)
  • Would kill commercial AU development

Option 5: Single Custom License​

Pros:

  • Tailored exactly to our needs

Cons:

  • Not OSI-approved
  • Legal review overhead
  • Confuses contributors
  • Reduces trust and adoption

Decision Rationale​

Dual MIT/Apache-2.0 chosen because:

  1. Rust ecosystem alignment: This is the standard for Rust projects (used by rustc, cargo, most popular crates)
  2. Maximum compatibility: Satisfies both GPLv2 and patent-protection needs
  3. Simple for contributors: Well-understood, no CLA required
  4. Commercial-friendly: Enables ecosystem growth and AU marketplace
  5. Patent protection available: Apache-2.0 option provides explicit patent grant
  6. Proven model: Thousands of successful Rust projects use this

Consequences​

Positive​

  • Broad adoption: No licensing barriers for users
  • Commercial-friendly: Companies can build proprietary AUs
  • Patent protection: Available via Apache-2.0 option
  • Rust ecosystem fit: Aligns with community norms
  • Simple contribution: No CLA, standard license headers
  • Model flexibility: Separate model licenses allow appropriate restrictions
  • Clear separation: Software (permissive) vs Models (potentially restricted)

Negative​

  • Two licenses to maintain: Slightly more overhead (mitigated by being standard)
  • License choice complexity: Users must choose (mitigated by "either" language)
  • Model license complexity: Different rules for software vs models (necessary trade-off)

Neutral​

  • No copyleft: Proprietary forks are possible (but we accept this trade-off)
  • Attribution requirement: Both licenses require attribution (standard practice)

Implementation Notes​

Repository Structure​

Each code repository gets:

repo/
├── LICENSE-MIT
├── LICENSE-APACHE
├── README.md (explains dual licensing)
└── src/
    └── lib.rs (standard license header)

License Headers​

Standard Rust license header in source files:

// Licensed under either of
//
// * Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
// * MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)
//
// at your option.

Contribution Guidelines​

Standard Rust contribution clause in CONTRIBUTING.md:

Unless you explicitly state otherwise, any contribution intentionally submitted
for inclusion in the work by you, as defined in the Apache-2.0 license, shall be
dual licensed as above, without any additional terms or conditions.

Model Licensing​

Repositories containing models:

agx-ocr/
├── LICENSE-MIT (for code)
├── LICENSE-APACHE (for code)
├── MODEL-LICENSE (for model weights)
├── models/
│   ├── deepseek-ocr/
│   │   ├── README.md (cites upstream DeepSeek license)
│   │   └── LICENSE (DeepSeek Model License)
│   └── tinker-delta-1.5b/ (future AGEniX-trained model)
│       ├── README.md (cites AGEniX Model License)
│       └── LICENSE (AGEniX Model License)

Model License Template (Future AGEniX-Trained Models)​

When we train our own models, use a permissive license with use-based restrictions:

Inspiration: DeepSeek Model License

Key terms:

  • Commercial use: Permitted
  • Modification: Permitted
  • Redistribution: Permitted with attribution
  • Patent grant: Included
  • Use restrictions: Prohibit illegal or harmful activities
  • Attribution: Required
  • Copyleft: No (derivatives can use different licenses)

README Licensing Section​

Each repo README.md should include:

## License

This project is dual-licensed under either of:

 * Apache License, Version 2.0, ([LICENSE-APACHE](LICENSE-APACHE) or http://www.apache.org/licenses/LICENSE-2.0)
 * MIT license ([LICENSE-MIT](LICENSE-MIT) or http://opensource.org/licenses/MIT)

at your option.

### AI Model Weights

AI model weights (if included) are licensed separately. See [MODEL-LICENSE](MODEL-LICENSE) for details.

### Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted
for inclusion in the work by you, as defined in the Apache-2.0 license, shall be
dual licensed as above, without any additional terms or conditions.

Model Distribution Strategy​

Bundled Models (e.g., agx-ocr + DeepSeek)​

Approach: Include model weights in repository with upstream license

Files:

  • Code: MIT/Apache-2.0
  • Model: DeepSeek Model License (preserved)
  • Clear README explaining license split

User agreement: Implied acceptance when downloading/using

Downloaded Models (Future)​

Approach: Prompt user to accept model license before download

Flow:

$ agx model install tinker-delta-1.5b
This model is licensed under the AGEniX Model License.
Key terms:
  - Commercial use permitted
  - Attribution required
  - Prohibited: illegal/harmful activities

View full license: https://agenix.sh/licenses/model
Accept? [y/N]: y
Downloading...

Storage:

  • Models cached in ~/.agx/models/
  • License stored alongside: ~/.agx/models/<model-name>/LICENSE
  • Contribution guidelines flow from this decision
  • AU marketplace licensing will build on this foundation
  • Future ADR: AGEniX Model License specification (when we train models)

References​