Qwen releases QwQ-32B, small reasoning model which rivals with DeepSeek-R1 and o1-mini.

TL;DR:

• Qwen is a serious player, just like DeepSeek
• a 32B model that rivals with a 671B mixture of experts (37B params activated)
• weights available on huggingface, no paper yet, you can chat with it on chat.qwen.ai
• they used 2 stages of reinforcement learning, 1st one for math and coding tasks and 2md one for general capabilities
• for math, reward is based on accuracy verifier, for coding it’s based on predefined tests cases that are actually executed
• let’s wait for other benchmarks, Qwen’s look decent but interested to see if it replicates / what it looks like in the open LLM leaderboard