Qwen 3.8 27B catches up to giant models but burns extreme compute to do it

Qwen 3.8 27B scored 52 on the Artificial Analysis Intelligence Index. That is the exact same score as the large GPT-5.6 Luna and just one point behind the 753-billion parameter GLM-5.2. For a model with only 27 billion parameters, this is an exceptional result.

The end of relying solely on parameters

These figures change the way developers measure the power of models. Until now, it was true that the quality of an answer depended directly on the volume of data learned in the parameters of the neural network.

Inference as a substitute for a giant brain

Qwen does not achieve these results because it is miraculously smarter at its core than the competition. It uses a reasoning mode that pushed the benchmark eval to generate 160 million tokens (the median for comparable models is 43 million). The model thinks long and hard before answering.

Speed and tokens complicate local hardware use

Even though Qwen looks like an ideal candidate for local execution on paper, its reasoning mode generates absurdly long chains of thought. While the small size pleases owners of smaller GPUs, the delivery speed of the final answer suffers significantly.

Real utility in code will be the deciding factor

A clean benchmark score is one thing, but practical usability as a coding agent will reveal the truth. If the model spends five minutes generating thoughts while writing a script and then gets stuck in a loop, its paper victory in the benchmark will collapse.

Lilith's verdict

Reading benchmarks is no longer enough. Qwen has shown that even a small model can reach the top of the chart if it gets unlimited time to overthink every little detail.