curator-research/Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-ON

TEXT GENERATIONPricing:Input $0.2 / Output $0.8Concurrent Unit Cost:2Model Size:30BQuant:FP8Context Size:32kPublished:Oct 1, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-ON is a 30 billion parameter experimental language model from Curator Research, built on the Nemotron-Cascade-2-30B-A3B architecture and trained on Llama 3.3 70B outputs. This "thinking-on" variant is specifically designed for structured extraction of entities, facts, and citations from documents, serving as a companion to its "thinking-off" sibling. It features a 32768 token context length and is optimized for enrichment tasks, with its held-out evaluation currently pending.

Loading preview...

Model Overview

Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-ON is a 30 billion parameter experimental release from Curator Research, utilizing the nvidia/Nemotron-Cascade-2-30B-A3B architecture. It is trained on outputs from Llama 3.3 70B, focusing on "thinking-on" capabilities for complex information extraction tasks. This model is designed for structured extraction of entities, facts, and citations from documents, with a reported validation loss of 0.0701 at its selected checkpoint.

Key Capabilities

  • Structured Information Extraction: Specialized in extracting entities, facts, and citations from documents in a structured format.
  • "Thinking-On" Mode: This variant is configured for a "thinking-on" approach, which can be enabled via the chat template's enable_thinking option, potentially allowing for more complex reasoning during extraction.
  • Llama 3.3 Lineage: Benefits from training on high-quality outputs generated by Llama 3.3 70B.

Performance and Usage

While held-out evaluation for this specific "thinking-on" model is pending, its "thinking-off" sibling, Llama-3.3-Nemotron-Cascade-2-30B-A3B-Curator-0.1.0-Enrichment-Thinking-OFF, demonstrated strong serving figures, including 154 tokens/s single-stream decoding on an A100 GPU. The model can be loaded using runtimes compatible with its base architecture and served with vLLM 0.29. Users should evaluate this specific revision on their own data before production use, as training data, prompts, and output schema remain private.

Finetunes
1 models