Proposals
This section contains proposals for significant KAITO features and model integrations. Some proposals describe how new OSS models are evaluated and integrated, while others define broader platform or API changes.
Proposal Template
Before creating a new proposal, use the appropriate template in docs/proposals.
Current Proposals
Below are the current proposals in various stages of integration:
Provisional Status
- Llama 3.3 70B Instruct - Meta's multilingual instruction-tuned 70B model
- Qwen2.5 Coder - Qwen2.5 series for code generation
- Phi-4 Instruct - Microsoft's latest Phi-4 instruction-tuned model
- Distributed Inference - Support for distributed inference across multiple GPUs
- Model as OCI Artifacts - Packaging models as OCI artifacts
- RAGEngine Guardrails UX and API - Minimal enable switch plus ConfigMap-based YAML policy for RAGEngine guardrails
Integrated Status
- Mistral Instruct - Mistral AI's instruction-tuned model
- Mistral - Base Mistral model
- Phi-2 - Microsoft's Phi-2 small language model
- Phi-3 Instruct - Microsoft's Phi-3 instruction-tuned model
Proposal Process
For detailed information about the model onboarding process, see the Model Onboarding Guide.
Step 1: Create a Proposal
Use the model proposal template to describe the target OSS model, including licensing, usage statistics, and technical requirements.
Step 2: Model Validation
KAITO maintainers validate and test the proposed model using the specified runtime.
Step 3: Image Publishing
If licensing allows, model images are published to Microsoft Container Registry (MCR).
Step 4: Integration
Implement preset configurations and inference interfaces for the model.
Step 5: Testing
Add comprehensive E2E tests to ensure the model works correctly with KAITO.
Contributing a Proposal
To contribute a new model proposal:
- Fork the KAITO repository
- Copy the model template to
docs/proposals/YYYYMMDD-<model-name>.md - Fill out all required sections
- Submit a pull request for review
The proposal status will be updated as it progresses through the integration pipeline.