Researchers just shared the latest Light-R1-32B AI model that’s optimized to solve complex and the most high-level math sums. The model is now up for grabs on Hugging Face under the permissive Apache 2.0 license.
This means it’s free of cost for businesses and any researchers that want to take, use, fine-tune, and edit as per their customized likings. The same is the case for commercial reasons. The 32 billion parameter model goes beyond the usual performance of similar sized models that are open source in design like DeepSeek’s R1-Distill-Llama-70B. These are on the benchmark designs that feature 15 math sums created for the most advanced students and have allotted timeframes of nearly 3 hours.
The model is created by a host of Chinese developers and so far, it’s been able to surpass previous models in the industry, which were designed on the most competitive math benchmark.
In this case, researchers were able to finish training of the system in less than six hours through the 12 Nvidia H800 GPUs for $1000. This makes this system one of the most easily accessed and practical systems for creating high performance math related AI models. It’s very important to realize that this model was trained using a version of Alibaba’s open source Qwen 2.5-32B Instruct. This alone is said to entail the highest upfront in regards to training costs.
Alongside this model, the team has rolled out training datasets and scripts as well as tools for evaluation. This provides the most transparent and easily accessible design system for creating AI models that are based on math. We can see how this latest launch follows in those footsteps as seen with rivals like Microsoft Orca-Math.
To help the system tackle difficult math and reasoning problems, the researchers carried out training on models that weren’t equipped with a COT of a long chain of thought for reasoning. Instead, they based the use of SFT or supervised fine-tuning and DPO or direct preference optimization to solve problems.
During the evaluation phase, the model surpasses previous systems for this purpose and achieved higher scores. So the improvement is major and worth a mention that curriculum-related training methods better mathematical reasoning, even when you don’t use specific models without COT.
To make sure the benchmarking was fair, researchers claim to have decontaminated all training materials against reasoning benchmarks. Therefore, experts claim businesses can benefit in many great ways.
This example makes it worthwhile and attractive for AI developers, companies, and software engineers searching to combine and customize the model for various proprietary uses.
To conclude, the researchers hail the model as a transparent, cost-effective, and optimized system for solving high-level math problems. The fact that it’s open source and goes beyond the standard benchmarks set by DeepSeek at a fraction of the training costs says it all. What do you think?
Image: DIW-Aigen
Read next: Meta Expands Its Use of Facial Recognition Tools To Combat Scams in the UK and EU
This means it’s free of cost for businesses and any researchers that want to take, use, fine-tune, and edit as per their customized likings. The same is the case for commercial reasons. The 32 billion parameter model goes beyond the usual performance of similar sized models that are open source in design like DeepSeek’s R1-Distill-Llama-70B. These are on the benchmark designs that feature 15 math sums created for the most advanced students and have allotted timeframes of nearly 3 hours.
The model is created by a host of Chinese developers and so far, it’s been able to surpass previous models in the industry, which were designed on the most competitive math benchmark.
In this case, researchers were able to finish training of the system in less than six hours through the 12 Nvidia H800 GPUs for $1000. This makes this system one of the most easily accessed and practical systems for creating high performance math related AI models. It’s very important to realize that this model was trained using a version of Alibaba’s open source Qwen 2.5-32B Instruct. This alone is said to entail the highest upfront in regards to training costs.
Alongside this model, the team has rolled out training datasets and scripts as well as tools for evaluation. This provides the most transparent and easily accessible design system for creating AI models that are based on math. We can see how this latest launch follows in those footsteps as seen with rivals like Microsoft Orca-Math.
To help the system tackle difficult math and reasoning problems, the researchers carried out training on models that weren’t equipped with a COT of a long chain of thought for reasoning. Instead, they based the use of SFT or supervised fine-tuning and DPO or direct preference optimization to solve problems.
During the evaluation phase, the model surpasses previous systems for this purpose and achieved higher scores. So the improvement is major and worth a mention that curriculum-related training methods better mathematical reasoning, even when you don’t use specific models without COT.
To make sure the benchmarking was fair, researchers claim to have decontaminated all training materials against reasoning benchmarks. Therefore, experts claim businesses can benefit in many great ways.
This example makes it worthwhile and attractive for AI developers, companies, and software engineers searching to combine and customize the model for various proprietary uses.
To conclude, the researchers hail the model as a transparent, cost-effective, and optimized system for solving high-level math problems. The fact that it’s open source and goes beyond the standard benchmarks set by DeepSeek at a fraction of the training costs says it all. What do you think?
Read next: Meta Expands Its Use of Facial Recognition Tools To Combat Scams in the UK and EU
