OpenThinker-7B
AI model · Language model (LLM) · Text → Text by Bespoke Labs
A decensored 7B reasoning model for straightforward and aligned AI responses.
OpenThinker-7B is a 7B parameter reasoning model from Bespoke Labs that offers decensored yet safe responses. It is designed for local execution via Ollama and Hugging Face, providing a straightforward alternative for complex reasoning tasks.
Model specs
The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.
- Launched
- 2025-01-28
- Languages
- English
Category: Generative AI
- Parameters
- 7B
- Sizes and variants
- OpenThinker-7B, OpenThinker-32B
- Access
- Download (Hugging Face, Ollama)
- Released
- 2025-01-28
- Parameter Count
- 7B
- Base Model
- Qwen2.5-7B-Instruct
- 7B parameter reasoning model
- Trained on OpenThoughts-114k dataset
- Decensored yet safety-aligned responses
Overview
OpenThinker-7B is a compact reasoning model developed by Bespoke Labs for researchers and developers seeking decensored AI capabilities. It is a 7B parameter model built by fine-tuning the Qwen2.5-7B-Instruct base model on the OpenThoughts-114k dataset, which contains math, coding, and graduate-level reasoning data. The model is primarily distributed as open weights, allowing for local execution via platforms like Ollama, where it has reached over half a million downloads. It serves as a key component of the DataComp and Bespoke Labs community effort to curate open reasoning datasets and models.
Reasoning and Alignment
- OpenThoughts-114k Training: The model is trained on the OpenThoughts-114k reasoning dataset, which consists of math, coding, and graduate-level questions and answers.
- Decensored Reasoning: The model appears to be decensored even though no specific fine-tuning was performed to remove political censoring.
- Safety Alignment: Despite being decensored, the model remains aligned and refuses to answer harmful or unsafe questions.
Deployment and Access
- Hugging Face Availability: The model is available for download on the Hugging Face platform.
- Ollama Integration: Users can run the model locally using the Ollama platform with the command "ollama run openthinker:7b".
- Community Adoption: The model has achieved approximately half a million downloads on the Ollama platform.
Use cases
- Researchers use the model to explore how fine-tuning on specific datasets affects model alignment and political censorship.
- Developers run the model locally via Ollama to perform complex reasoning tasks without relying on cloud-based censored APIs.
- Data scientists utilize the OpenThoughts-114k dataset foundations to build and test specialized reasoning applications.
Specifications
- Parameters
- 7B
- Sizes and variants
- OpenThinker-7B, OpenThinker-32B
- Access
- Download (Hugging Face, Ollama)
- Released
- 2025-01-28
- Parameter Count
- 7B
- Base Model
- Qwen2.5-7B-Instruct
- Training Dataset
- OpenThoughts-114k
- Ollama Command
- ollama run openthinker:7b
- License Type
- Permissible (MIT for training data source)
How it works
- Model Training: The model was developed by fine-tuning a Qwen2.5-7B-Instruct base model on data generated by DeepSeek-R1.
- Dataset Integration: The training utilized the OpenThoughts-114k dataset, a curated collection of math, coding, and graduate-level reasoning tasks.
- Distribution: Users download the model weights from Hugging Face or pull the model through the Ollama library.
- Local Execution: The model is executed locally or in a private environment to perform complex reasoning tasks without political censoring.
FAQ
How can I access OpenThinker-7B?
The model is available for download on Hugging Face and can be run locally via Ollama using the command "ollama run openthinker:7b".
What data was used to train OpenThinker-7B?
The model was trained on the OpenThoughts-114k dataset, which includes math, coding, and graduate-level questions, using DeepSeek-R1's data applied to a Qwen base model.
Is OpenThinker-7B safe to use?
Yes, while the model is decensored regarding political views, it is still aligned to refuse harmful and unsafe questions.
Added by AntSpace from public sources (bespokelabs.ai).