Radar ·
Layer pruning of Qwen3-8B-Base with LaCo reduces model size by 16.7%
A post reports removing up to 10 layers from Qwen3-8B-Base with LaCo. With 6 layers removed, the model is 16.7% smaller, keeping 78% to 94% of reasoning scores with a perplexity ratio of 1.68x, according to the author.
According to the post, the author removed up to 10 layers from Qwen3-8B-Base using LaCo (Layer Collapse). With 6 layers removed, the model is 16.7% smaller, and the author says it keeps between 78% and 94% of reasoning scores, with a perplexity ratio of 1.68x. The report shows a trade-off between layer pruning, model size and quality, useful for thinking about cheaper inference.
This report does not describe a Rota Nacional feature. To study the material with AI, do not paste personal data, CPF, CNPJ, e-mail addresses or credentials into external tools. To verify metrics, benchmarks and experimental conditions, consult the original post, since the Radar summary does not include those details.
Original source: Poda de camadas do Qwen3-8B-Base com LaCo para reduzir o tamanho do modelo