Uploaded August 2025 | Updated September 2026, 1 week ago
Generative AI often forces a choice: creative outputs or coherent ones. Min-p sampling is a new approach that lets you have both.
Min-p sampling is a new technique that makes large language models more creative without sacrificing structure — and it’s already reshaping the open-source AI ecosystem.
Learn about Min-p sampling in this video as Allen Roush, Lead AI researcher, explains how it differs from top-k and top-p, and shows why it’s becoming the go-to approach for open-source developers.
You’ll see how Min-p empowers higher temperature sampling, enabling more diverse, less repetitive, and still coherent outputs.
Key takeaways:
- Why traditional sampling (greedy, top-k, top-p) falls short at high temperatures
- How Min-p dynamically balances creativity and reliability
- Adoption stories from Hugging Face, VLLM, and other open-source frameworks
- Practical demos running Min-p locally
Who should watch:
- AI researchers and engineers working with open-source LLMs
- Developers interested in improving generative AI creativity
- Teams building applications that need diverse but controlled AI outputs
----
Follow Thoughtworks on YouTube for more deep dives into AI research, engineering, and innovation.
Read about our AI software engineering capabilities and offerings here: thoughtworks.com/en-us/what-we-do/ai/ai-enabled-software-engineering
Generative AI often forces a choice: creative outputs or coherent ones. Min-p sampling is a new approach that lets you have both.
Min-p sampling is a new technique that makes large language models more creative without sacrificing structure — and it’s already reshaping the open-source AI ecosystem.
Learn about Min-p sampling in this video as Allen Roush, Lead AI researcher, explains how it differs from top-k and top-p, and shows why it’s becoming the go-to approach for open-source developers.
You’ll see how Min-p empowers higher temperature sampling, enabling more diverse, less repetitive, and still coherent outputs.
Key takeaways:
- Why traditional sampling (greedy, top-k, top-p) falls short at high temperatures
- How Min-p dynamically balances creativity and reliability
- Adoption stories from Hugging Face, VLLM, and other open-source frameworks
- Practical demos running Min-p locally
Who should watch:
- AI researchers and engineers working with open-source LLMs
- Developers interested in improving generative AI creativity
- Teams building applications that need diverse but controlled AI outputs
----
Follow Thoughtworks on YouTube for more deep dives into AI research, engineering, and innovation.
Read about our AI software engineering capabilities and offerings here: thoughtworks.com/en-us/what-we-do/ai/ai-enabled-software-engineering
