Integrate NVIDIA NIM as an Inference Provider
Resolved 💬 6 comments Opened Apr 23, 2026 by HackWidMaddy Closed Apr 23, 2026
💡 Likely answer: A maintainer (etraut-openai, contributor)
responded on this thread — see the highlighted reply below.
What variant of Codex are you using?
CLI / App (works for both)
What feature would you like to see?
I would like Codex to support NVIDIA NIM as an official inference provider.
NVIDIA NIM provides optimized, production-ready inference microservices for running LLMs and other AI models efficiently on NVIDIA GPUs. Integrating NIM into Codex would:
- Enable high-performance local and on-prem inference
- Provide better GPU utilization and lower latency
- Allow developers to use enterprise-grade AI deployment pipelines
- Expand support beyond current providers to include NVIDIA’s ecosystem
This could be implemented as:
- A selectable inference backend (similar to existing providers)
- Configuration support for NIM endpoints (local or remote)
- Optional GPU-aware optimization settings
Additional information
NIM is becoming a key component in enterprise AI stacks, especially for teams deploying models on private infrastructure. Supporting it in Codex would make the tool more flexible for advanced users, researchers, and organizations working with NVIDIA hardware.
Potential benefits:
- Better alignment with modern AI infra (DGX, on-prem clusters, etc.)
- Strong appeal for developers already using NVIDIA AI Enterprise
- Improved performance for coding-heavy workloads
6 Comments
Thanks for the suggestion, but we're unlikely to add this.
HOW TO DO THIS
Since NVIDIA NIM uses a different API schema and not a fully OpenAI-compatible one, the practical approach would be through an adapter/proxy layer.
Codex -> Adapter Proxy -> NVIDIA NIM
The proxy would translate OpenAI-style
/v1/chat/completionsrequests into NIM’s expected request format and then convert the response back for Codex.This would allow NIM integration even without native provider support, while still keeping the developer workflow smooth for users running local or enterprise NVIDIA infrastructure.
how to do it
I am working on it will share repo of it soon
Guess What its out now: https://github.com/HackWidMaddy/OpenCodex