Qwen3 is one of the most capable open-weights models available, yet no feature-level interpretability tools existed for it, until now. BluelightAI’s cross-layer transcoders open Qwen3’s internal representations to inspection for the first time.

Cross-layer transcoders (CLTs) are AI models trained to interpret another AI model. They break down the input of each layer of the source model into combinations of interpretable features, and use those features to reconstruct the output of that layer and subsequent layers. This helps explain the work that each layer performs in terms of human-interpretable concepts.

CLTs were developed by researchers at Anthropic to make discoveries about how LLMs perform tasks, and recently open source CLTs have been released for Gemma 2 and Llama 3.2. However, no CLTs have been available for the Qwen3 models, which are among the most popular open-weights models today. We're fixing that problem with these Qwen3 CLTs.

We can use a CLT to give an LLM a "brain scan" as it processes an input, and identify which features are activated—in other words, what the model is thinking about. In some scenarios, knowing the features that are active can be enough to let you predict what the model will or should do.

We can go even deeper, though. Researchers have used CLTs to build attribution graphs for prompts, which trace the model's computation through individual features to understand the factors that influence its output. This can produce very detailed explanations of how a model is able to perform a particular task.

We used Cobalt to construct topological graphs from the features from each model. These graphs are multiresolution representations of the set of features, displaying groups of related features as nodes with connections between them. These graphs are available to browse in the feature dashboard.

Graphs are built based on different similarity measures for each feature. We built graphs based on feature encoder vectors, feature decoder vectors, and feature coactivation patterns. For encoder and decoder vectors, we built one graph for each layer, as it is not immediately obvious that feature vectors from different layers are comparable.

AI developers, researchers, and teams building production models who want deeper understanding of how Qwen3 transforms inputs and where failure modes arise.