<RETURN_TO_BASE

NVIDIA Releases Open-Source Open Code Reasoning Models with Unmatched Code Intelligence

NVIDIA has released its Open Code Reasoning models (32B, 14B, 7B) as open-source under Apache 2.0, delivering top-tier performance in code reasoning tasks and broad compatibility with popular AI frameworks.

NVIDIA's Open Code Reasoning Models

NVIDIA has open-sourced its Open Code Reasoning (OCR) model suite, featuring three large language models with 32B, 14B, and 7B parameters. These models are designed specifically for advanced code reasoning and problem-solving tasks and are available under the Apache 2.0 license.

Benchmark Performance

The OCR models demonstrate superior performance compared to OpenAI’s o3-Mini and o1 (low) models on the LiveCodeBench benchmark, which evaluates debugging, code generation, and logic completion. The 32B OCR model leads the leaderboard for reasoning capabilities among open models.

This success is attributed to both the model architecture and NVIDIA’s specialized OCR dataset, a high-quality, code-focused training corpus emphasizing instruction following, reasoning, and multi-step code problem solving. The dataset enables a 30% improvement in token efficiency, allowing the models to generate accurate code and logical outputs with fewer tokens.

Model Variants for Different Needs

The suite includes:

  • OpenCodeReasoning-Nemotron-32B: High-performance inference and research
  • OpenCodeReasoning-Nemotron-14B: Balanced reasoning capability and compute efficiency
  • OpenCodeReasoning-Nemotron-7B: Suitable for resource-constrained environments with competitive benchmarks

All models utilize NVIDIA’s Nemotron architecture, a transformer-based backbone optimized for multilingual, multi-task learning. Model weights and configurations are accessible on Hugging Face.

Compatibility with Popular Inference Frameworks

The models support out-of-the-box integration with:

  • llama.cpp for lightweight CPU/GPU inference
  • vLLM for optimized GPU serving and speculative decoding
  • Transformers by Hugging Face for training and evaluation
  • TGI (Text Generation Inference) for scalable API deployment

This ensures easy adoption by developers, researchers, and enterprises with existing AI infrastructure.

Advancing Open Code Intelligence

By open-sourcing these high-performance code reasoning models, NVIDIA significantly contributes to the open AI ecosystem, providing robust alternatives to proprietary solutions. This enables the community to build, fine-tune, and deploy sophisticated reasoning models for developer copilots, automated code review, and code generation services.

Explore the 32B, 14B, 7B models and the 32B instruction-tuned variant to experience cutting-edge open code intelligence.

🇷🇺

Сменить язык

Читать эту статью на русском

Переключить на Русский