J-lens Qwen

jlens-qwen36 is a visual debugger for local Qwen3.6-27B (4-bit) inference on Apple Silicon. It installs locally, loads the model through MLX, fits a Jacobian lens to the model’s residual stream, and presents a web-based layer-wise token prediction view. Each token position at each layer gets a ranked readout of what the model is predicting — not just the final output logits, but the intermediate representations at every depth. The workflow is local-first: install via the provided Python tooling, point it at your model weights, prompt it, and watch the visualization update in the browser.

The differentiating feature is the Jacobian lens itself, ported from research on verbalizable representations (the vendor cites Anthropic’s 2025 work on global workspace theory in LLMs). A Jacobian lens projects intermediate hidden states back into vocabulary space using the model’s own output matrix plus a fitted linear correction, making per-layer token predictions interpretable without requiring a separate probe model. This is the mechanism that lets you see a model writing ‘I will comply’ in its output while its mid-layer representations are predicting something structurally different — the kind of signal that post-hoc logit analysis does not expose.

The tool fits researchers and developers doing mechanistic interpretability work on this specific model and hardware stack. It covers four concrete tasks: inspecting latent concepts mid-generation, debugging refusal or suppressed reasoning, comparing per-layer token predictions visually, and fitting and testing custom Jacobian lenses. Where it breaks is equally specific: it is architected around Qwen3.6-27B (4-bit) and MLX. There is no documented path to other model families, no API for programmatic access, and no support for CUDA or non-Apple hardware. A read-only browser demo exists at jlens.wezzard.com for evaluation without local installation.

The project is Apache-2.0 licensed, fully open-source, and self-hosted. The Python environment is managed via uv, and the repository includes scripts, tests, and a web directory for the visualization frontend. No external data is transmitted; all inference and lens-fitting runs locally on the user’s machine.