QwQ-32B, from Alibaba Qwen team, is a new open-source 32B LLM achieving DeepSeek-R1 level reasoning via scaled Reinforcement Learning. Features a "thinking mode" for complex tasks.
Check out QwQ-32B, a new open-source language model from the Qwen team. It's achieving something remarkable: reasoning performance comparable to DeepSeek-R1, but with a model that's 20 times smaller (32B parameters vs. 671B)!
This is a big deal because:
🤯 Size/Performance Ratio: It punches way above its weight class in reasoning, math, and coding tasks. 🧠Scaled Reinforcement Learning: They achieved this by scaling up Reinforcement Learning (RL) on a strong foundation model (Qwen2.5-32B). 🤔 "Thinking Mode": Like some other recent models, it has a special "thinking mode" (activated with tags) that allows for longer chains of thought. ✅ Open Source: Model weights are available under Apache 2.0.
The implications of getting this level of reasoning performance from a 32B model are huge. It opens up possibilities for deploying powerful AI on less powerful hardware, reducing costs, and making advanced reasoning capabilities more accessible.
This model is blazing fast and available for Pro users in beta as of yesterday (3.6.25). Free of censorship, and system prompts are editable without your data going to any foreign government
Hi everyone!
Check out QwQ-32B, a new open-source language model from the Qwen team. It's achieving something remarkable: reasoning performance comparable to DeepSeek-R1, but with a model that's 20 times smaller (32B parameters vs. 671B)!
This is a big deal because:
🤯 Size/Performance Ratio: It punches way above its weight class in reasoning, math, and coding tasks.
🧠Scaled Reinforcement Learning: They achieved this by scaling up Reinforcement Learning (RL) on a strong foundation model (Qwen2.5-32B).
🤔 "Thinking Mode": Like some other recent models, it has a special "thinking mode" (activated with tags) that allows for longer chains of thought.
✅ Open Source: Model weights are available under Apache 2.0.
Available now in Qwen Chat and HF Spaces.
The implications of getting this level of reasoning performance from a 32B model are huge. It opens up possibilities for deploying powerful AI on less powerful hardware, reducing costs, and making advanced reasoning capabilities more accessible.
@sonu_goswami2 It reveals all the details of its reasoning, clear and transparent – just like R1!
Grimo
@zaczuo cmon this is so fckin good! we've been playing with it on qwen chat as well as groq cloud, performance was astonishing! great job qwen team!
@sonu_goswami2 @zaczuo Nice! Will check it out!
I use hugging face every day for my job. I didn't know that models there can be launched on PH. Very cool!
MGX (MetaGPT X)
This score and size are amazing, I have a few questions
What is the difference between it and DeepSeek R1?
Do you have plans to make a larger size reasoning model?
Thank you for your contribution to the open source community!