BlenderBot 3
A 175B open-domain chatbot with internet access and long-term memory
BlenderBot 3 is a 175-billion-parameter dialogue model for open-domain conversation, built by Meta AI researchers and described in an August 2022 technical report. It was deployed on a public web page to talk with real users, with weights and code released so the community could study agents that keep learning from interaction.
- 175B parameter dialogue model for open-domain conversation
- Deployed on a public web page for organic users
- Weights and code released publicly
- Internet access and long-term memory built in
- Technical report posted 5 August 2022
What it is
BlenderBot 3 is a dialogue model built for open-domain conversation, meaning chat on any topic rather than a fixed task. It has 175 billion parameters and can reach the internet and keep a long-term memory of earlier exchanges. The work comes from a team of Meta AI researchers, who published a technical report on it in August 2022. The model was trained on many user-defined tasks. Its authors released both the weights and the code, and put the agent on a public web page so ordinary users could talk to it. Their stated aim was to let the research community study agents that improve through use while staying responsible.
How it came about
The technical report was first posted on 5 August 2022 and revised twice within a week, with the final version dated 10 August 2022. Kurt Shuster led a group of eighteen authors including Jason Weston, Stephen Roller and Y-Lan Boureau. BlenderBot 3 follows earlier open-domain agents from the same group, cited in the report as Roller et al. 2021 and Komeili et al. 2022. Alongside the model, the team set out a plan for continual learning from the conversations gathered during public deployment, with that data also intended for public release.
Inside the model
The report describes the model as a 175B parameter dialogue system, and covers its architecture, its training scheme and the tasks it was trained on. Rather than answering from fixed knowledge alone, the agent can search the internet and hold a long-term memory, which lets it bring outside information into a conversation and refer back to earlier turns. Training used a large set of user-defined tasks. The authors also detail the safety mechanisms built into the public deployment. Full layer counts, tokenizer details and training compute are not given in the summary provided here.
What it can do
Human evaluations reported in the technical paper found BlenderBot 3 better than existing open-domain dialogue agents, including the group's own earlier models. Its distinguishing features are open-domain chat, live internet access and a long-term memory. The report frames the work as a research programme for agents that keep improving through interaction with users, rather than as a fixed release.
How to use it
Both the model weights and the code were released publicly. The agent was also deployed on a public web page where organic users could converse with it. The report states that data collected during deployment would be publicly released as part of the continual learning plan.
Safety and testing
The technical report devotes attention to the safety mechanisms used in the public deployment. The stated goal of the research programme is agents that engage responsibly and learn through interaction.
Reception and debate
Human evaluations in the report compare BlenderBot 3 with prior open-domain dialogue systems and rate it higher. Beyond those reported comparisons, the summary provided here gives no further reception detail.
Why it matters
The authors position the release as a way for the community to study agents that improve over time through real user interaction. Releasing weights, code and deployment data was intended to support that study of continually learning, responsible conversational agents.
Questions people ask about BlenderBot 3
What is BlenderBot 3?
It is a 175-billion-parameter conversational model designed for open-domain chat, meaning it can discuss any topic. It can search the internet and keep a long-term memory of past exchanges. Meta AI researchers built it and published a technical report in August 2022.
How can I use it?
The authors released both the model weights and the code, so it can be downloaded and run. They also put the agent on a public web page where users could talk to it directly.
What makes it different from earlier chatbots?
It combines open-domain conversation with live internet access and a long-term memory. Human evaluations in the report rated it above existing open-domain dialogue agents, including the group's own earlier models.
What is the continual learning plan?
The team planned to use conversations collected from the public deployment to keep improving the model. That data was also intended to be released publicly so the community could study agents that learn through interaction.
Good reading and sources
There is no English Wikipedia article on BlenderBot 3 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2208.03188. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.