Hello NVIDIA Team,
I would like to request a rate limit increase for my NVIDIA NIM API account.
Current limit: 40 RPM
Requested limit: 200 RPM
Acceptable fallback: 100–150 RPM if 200 RPM is not available
Use case: personal / development-stage AI infrastructure tooling, not public production traffic.
I am developing an All-in-One AI Infrastructure Manager for Windows. The goal of the project is to make it easier to launch, configure, and manage local and remote AI infrastructure in a modular way.
The main research and engineering goal of this project is to design infrastructure that can significantly improve the practical usefulness of smaller and mid-sized models by compensating for their weaknesses with a well-structured knowledge system, memory, RAG, project-specific context, rules, and tool orchestration.
Instead of relying only on larger frontier models, I am building a modular system where model capability can be amplified through better surrounding infrastructure:
-
RAG service management
-
Memory service management
-
MCP tool management per project
-
AI skill/module management
-
Global shared rule systems for agents
-
Per-project AI configuration
-
Agent workflow orchestration
-
Local service lifecycle management on Windows
-
Knowledge and context routing for different AI tasks
-
Testing different model backends for coding, planning, reasoning, and tool-use workflows
I use NVIDIA NIM as one of the main model backends for development and evaluation. One of my goals is to make NVIDIA NIM a first-class backend option inside this AI infrastructure manager, alongside other OpenAI-compatible providers. A higher development limit would help me test NIM more realistically in agentic workflows.
The models I am testing include:
-
moonshotai/kimi-k2.5
-
z-ai/glm-5.1
-
qwen/qwen3-coder-480b-a35b-instruct
-
openai/gpt-oss-120b
-
nvidia/nemotron-3-super-120b-a12b
Why the current 40 RPM limit is restrictive:
My workflow is agentic and tool-based. A single development task may require multiple model calls for planning, tool selection, project context inspection, RAG/memory interaction, MCP tool configuration, debugging service startup, validating configuration changes, and analyzing logs.
Even with a single developer and conservative usage, 40 RPM can be reached during normal testing because the system itself is designed around multi-step AI infrastructure orchestration rather than one-off chat completions.
This is not a public service, not customer-facing traffic, and not batch scraping. It is development and evaluation usage for an AI infrastructure management tool.
Current safeguards and mitigations:
-
Single-user development usage
-
Local throttling
-
One active agent workflow at a time
-
Conservative retries and backoff
-
No public API traffic
-
No abusive parallel workloads
-
Request budgeting in agent instructions
-
Preference for batching context gathering where possible
A higher limit would make NVIDIA NIM practical as a development backend for testing real agentic AI infrastructure workflows under Windows. The current 40 RPM limit is usable for simple chat, but restrictive for multi-step AI infrastructure orchestration, especially when testing RAG, memory, MCP tools, per-project agent configurations, and model capability amplification through external knowledge and tooling.
Thank you for considering my request.