Congratulations on this release!
Hi Qwen team,
thank you again for your contributions to open source! This is a super innovative model, not only does it deliver better attention for maximum efficiency using the Qwen Next architecture, it also is native multimodal and has optional thinking onboard, so you no longer need seperate models. This is the first major step towards the future of AI. Love to see llama.cpp support at day 0 too!
Benchmarks look good as well, huge step up. I have tested it on Qwen Chat for a little bit and noticed improvements in general knowledge already, which was much needed, but I need to test it for far longer to make a decisive conclusion.
For now, let's party! Can't wait for the other sizes.
+1!
Thank you for all your epic work! It's really appreciated <3
This is the first open-weights model that passes all of my private personal trick questions benchmark, and the first open-weights model that is actually useful in Open Interpreter. I'm impressed!
My only complaint is that it does tend to make bad tool calls frequently, either calling tools that don't exist, or writing the tool call as plain text within a message. Open Interpreter is not the environment it was trained on though, so it's understandable it would become confused by the different types of "tools" it can use.