QWQ 32B REASONING MODEL

Summary from the TLDR AI newsletter

The Qwen team has trained an open weight, Apache 2.0, licensed model that performs on par with DeepSeek R1 and better than many of the larger distill models. They found that by building on outcome based rewards with formal verification and test-case checks, the model can continuously improve at math and code. Also, by mixing in general instruction following data later in RL training, the model can still align with human preference.