NVIDIA Open-Sources Open Code Reasoning Models (32B, 14B, 7B)

NVIDIA continues to advance the limits of developing open intelligence through open outsourcing Open symbol thinking suite (OCR) The three high -performance language models are designed for this purpose in order to think about the code and solve problems. 32B, 14B and 7B variables, all of them were released below Apache 2.0 license.

Summer to overcome the best

the Open symbol thinking (OCR) Models come with Note standard achievementsOutperform performance Openai’s O3-MINI and O1 (low) Models on LiveCOOOOOBENCH standard. LiveCodebeench is a comprehensive evaluation suite for thinking tasks of code such as correcting errors, generating code and completing logic in the environments of developers in the real world. In the direct comparison, the NVIDIA 32B OCR model tops the leading plate of the possibility of thinking about open models.

This jump is attributed to performance not only to the design of architecture, but to Nvidia’s Custom, “OCR Data Group” A high -quality training group centered around the code is designed to emphasize problem solving in instructions instructions, thinking and solving the multi -step code. According to NVIDIA, this leads to a 30 % improve the efficiency of the distinctive symbolAllow the models to produce a fine symbol and logical outputs with fewer symbols.

A typical assortment for each use case

The open symbol thinking comes Three teacher schedules:

• OpenCoderoning-Sunotron-32B
• OpenCodreasoning-Sunotron-14B
• OpenCodresoning-Sunotron-7B

Each scale balance model with performance. 32B variable provides the latest performance results for high performance inference and research; The 14B model provides strong thinking capabilities with low calculation requirements, the perfect 7B variable for resource restrictions while maintaining competitive performance on standards.

All models are trained using Nemotron architectureNVIDIA -based spine based on multi -language learning, multi -task. Obligations and typical configurations are available in embrace:

Compatible with open ecosystems for reasoning

The main feature of these models Compatibility outside the box With popular reasoning frameworks:

• llama.cpp For the conclusion of the CPU/light graphics processing unit
• vLLM To serve the improved GPU and dismantle speculation
• Transformers By embracing the training and evaluation pipelines
• TGI (Text generation conclusion) for the deployment of the API developmental API

This flexibility allows developers, researchers and institutions to connect these models in the infrastructure to maintain the current code with the least amount of public expenditures.

A step forward for the intelligence of the open symbol

With this version, NVIDIA contributes significantly to the increasing ecosystem of open code models. By targeting Code -A dominant field dominated by royal models historically-and is issued under an open and completely tolerant license. NVIDIA enables the AI ​​community and the broader developer of building advanced thinking models and spreading them in production.

The open symbol thinking adds to the growing NVIDIA wallet from the open LLMS and enhances its position on the development of an artificial intelligence that can be reached. Whether you are building joint developers, automatic code, or code generating services, these models offer a high -performance and effective alternative and a friend of society for closed solutions.

Check the 32B model, 14B model, 7B model and 32B instructions variable. Also, do not forget to follow us twitter.

Here is a brief overview of what we build in Marktechpost: