Qwen 3: The new open standard
Importance: 3 | # | qwen, opensource, nathan-lambert
The Qwen3 models’ scores are so good that we now get to see the conversation unfold as to whether Qwen’s models have the character and staying power of DeepSeek R1 and other recent entrants in the frontier model club. This is a major achievement — one that is downstream of substantial compute and personnel investment.
Qwen released 6 instruct models along with their base versions and quantized variants, including two sparse mixture of expert (MoE) models Qwen3-235B-A22B and Qwen3-30B-A3B1 that both have similar sparsity factors to the DeepSeek MoE architecture and 6 dense models 32B, 14B, 8B, 4B, 1.7B, and 0.6B parameters — all models with the Apache 2.0 License. For reference, the smallest Llama 4 model still has 109B total parameters and the 32B range is known to be very popular with open model users!
They've also released the base models!! The range of model sizes is incredible. 0.6B is incredibly cohearant for its size. I will wait for interesting things the community gets up to.