ckg-nvidia-nemotron 0.2.0

Compressed Knowledge Graph for NVIDIA Nemotron 3 Ultra 550B A55B — 71 nodes, 120+ edges. Architecture, training pipeline, reasoning modes, deployment, benchmarks, and agentic integrations.

ckg-nvidia-nemotron

Compressed Knowledge Graph · NVIDIA Nemotron 3 Ultra 550B A55B

71 nodes · 120+ edges · architecture, training pipeline, reasoning modes, deployment, benchmarks, agentic integrations.

Traversable via MCP — agents query declared relationships instead of inferring them.

pip install ckg-nvidia-nemotron

What's in the graph

Category	Concepts
Architecture	Hybrid Mamba-Transformer MoE, LatentMoE, MTP layers, 1M context
Training	20T tokens (NVFP4), 7 stages: Pretrain → SFT → RLVR → MOPD → MTP Boosting
RLVR	DAPO (NeMo RL), GRPO (NeMo Gym), 6 reward domains
MOPD	10+ teacher models, 2 iterative distillation rounds
Reasoning	Off / Regular / Medium + Reasoning Budget Control
Deployment	vLLM, SGLang, TensorRT-LLM, NIM Microservices
Hardware	DGX Cluster (4-node), GB300 DGX Station (selective expert offloading)
Parallelism	TP=2, PP=6/12, EP=32
Fine-tuning	LoRA via NeMo AutoModel, Megatron-Bridge, Unsloth
Agentic	Hermes Agent, OpenClaw, tool use, multi-step planning, deep research
Benchmarks	Intelligence Index 47.7, IFBench 81.4%, Ruler 95%, PinchBench 91%, Terminal-Bench 54%
Pricing	$0.50/$2.20 per 1M tokens (NIM API)

MCP tools

• search_concepts(query) — keyword search across all 70 nodes
• query_ckg(concept, depth) — subgraph traversal (1–3 hops)
• get_prerequisites(concept) — full upstream dependency chain
• list_domains() — package metadata
• get_benchmark_summary() — scores, competitive positioning, pricing
• get_deployment_guide(target) — nim / vllm / sglang / tensorrt / fine-tuning

Usage

Claude Desktop / Cursor / Claude Code

{
  "mcpServers": {
    "ckg-nvidia-nemotron": {
      "command": "uvx",
      "args": ["ckg-nvidia-nemotron"]
    }
  }
}

Python

from ckg_nvidia_nemotron.server import search_concepts, query_ckg, get_prerequisites

# Find the LatentMoE routing architecture
print(search_concepts("LatentMoE"))

# Traverse training pipeline
print(query_ckg("MOPD", depth=2))

# What must be in place before reasoning modes work?
print(get_prerequisites("Reasoning Mode Medium"))

HTTP server

ckg-nvidia-nemotron --port 8000
# MCP endpoint: POST http://localhost:8000/mcp

Benchmark

Grounded in the Compact Knowledge Graph benchmark (v0.6.2).

EVAL

benchmark: ckg-benchmark v0.6.2
dataset: huggingface.co/datasets/danyarm/ckg-benchmark
benchmarked: false
rag_baseline_f1: 0.123
graphrag_baseline_f1: 0.120
mean_tokens: 269
paper: github.com/Yarmoluk/ckg-benchmark/blob/main/paper/main.pdf

META

version: 0.1.1
domain: nvidia-nemotron
nodes: 71
edges: 120+
model_covered: Nemotron 3 Ultra 550B A55B
nvidia_release_date: 2026-06-04
license_of_model: OpenMDW-1.1
package_license: MIT
publisher: Graphify.md
publisher_url: https://graphifymd.com/pro/

---

Built by Graphify.md · Benchmark paper · patent pending