OpenThinker (Fully Tested): This NEW REASONING MODEL is QUITE CRAZY!

Summary

In this video, the creator introduces Open Thinker, an open-source AI model designed to push the limits of reasoning tasks. Available in 7B and 32B versions, it outperforms other open-data models in areas like math, code, and science. The 32B model, trained on a dataset of 114k tokens, shows impressive results, answering complex questions accurately. However, it struggles with coding tasks. Open Thinker is lauded for its potential as a proof of concept for open-source AI and local deployment on high-end hardware like RTX 4090. The video also highlights Ninja Chat, an AI platform with multiple models and tools, offering great value to users.

Takeaways

• 😀 Open Thinker is a state-of-the-art open-source AI model designed for reasoning tasks, with both model weights and datasets publicly available.
• 🤖 The model comes in two variants: a 7B and a 32B version, both trained on the Open Thoughts dataset with 173k questions.
• 📊 Open Thinker has been shown to outperform existing open-data reasoning models in several benchmarks, including math, code, and science.
• 📈 The 32B variant performs particularly well in reasoning tasks but can be run locally on powerful hardware like an RTX 4090.
• ⚙️ Open Thinker is built using a distillation approach and makes use of a verified dataset to improve performance over the unverified version.
• 🧠 The model’s reasoning abilities are particularly strong in questions involving logic, language, and math, but weaker in coding tasks.
• 📉 The 32B model outperforms the 7B version in most tasks but has a higher computational cost.
• 🔍 Open Thinker is the first model under 70B parameters to correctly answer complex logic questions, showing promise in open-source AI research.
• 💡 The project is designed to be a proof of concept for using data distillation in AI models, encouraging further development and fine-tuning by researchers.
• 🔧 Despite its strengths in reasoning, Open Thinker struggles with generating well-formatted code, an area for future improvement in the model’s capabilities.
• 🎯 Open-source datasets and models like Open Thinker allow researchers to make significant strides in AI development, providing opportunities to build more powerful models in the future.

Q & A

What is Open Thinker and what makes it unique?

Open Thinker is an open-source reasoning model that offers both model weights and training code. It is unique because it demonstrates the power of scaling data, verifying reasoning traces, and scaling model size to create a powerful reasoning model.

What are the two available sizes of the Open Thinker model?

The Open Thinker model comes in two sizes: a 7B variant and a 32B variant.

How does Open Thinker perform compared to other open-source reasoning models?

Open Thinker outperforms other open-source reasoning models on a range of reasoning benchmarks, including math, code, and science, although the 32B variant performs better in some cases compared to the 7B variant.

What is the dataset used to train Open Thinker, and how was it collected?

Open Thinker was trained using the Open Thoughts 114k dataset. The data was collected by gathering reasoning traces and solution attempts for 173k questions, which were then verified for correctness.

How do the 7B and 32B variants of Open Thinker differ in terms of performance?

The 32B model generally performs better, solving more complex tasks accurately, such as passing certain reasoning questions that the 7B model cannot. However, the 7B variant is also capable of performing well in various benchmarks.

Which model is recommended for users who want to run Open Thinker locally?

The 32B variant can be run locally on high-performance hardware like an RTX 4090, while the 7B variant can also be run but with potentially less optimal performance.

What are some challenges in using the Open Thinker models?

One challenge with the Open Thinker models is their performance in coding tasks, where both variants struggle to provide correctly formatted code or solve coding problems accurately.

What is the significance of the open data approach used in Open Thinker?

The open data approach used in Open Thinker allows other researchers to improve the dataset and potentially create even better models, making it easier for the AI research community to advance.

How does Open Thinker's 32B model compare to models with larger sizes, like the 70B model?

Despite being smaller, the 32B model performs similarly to larger models, such as the Deep Seek R1, and even outperforms them in certain benchmarks. It is also more resource-efficient and easier to run locally.

What additional features does Ninja Chat offer to users?

Ninja Chat provides access to over 10 different AI models, including GPT-4, Claude 3.5, and image generation models, all in one platform. It also includes tools for code generation, preview, and running Python code, which can be used to create and share charts and other outputs.