Get This Tool
J-lens Qwen
Pricing
- Model
- Free
Summary
Black-box LLM behavior is easy to rationalize after the fact and nearly impossible to inspect in the moment — especially when a model outputs one thing while its internal representations are pointing somewhere else entirely.
jlens-qwen36 fits a Jacobian lens to a local Qwen3.6-27B (4-bit) model running on Apple Silicon via MLX, then renders a layer-by-layer visualization of which tokens the model is pushing toward at each position during generation. The canonical demo is a blackmail email prompt: the model outputs a calm, compliant reply, but the workspace band exposes what the intermediate layers were actually predicting — a gap that standard logit inspection never surfaces. This is a single-model, single-hardware tool. It runs on macOS with MLX; there is no cloud backend, no REST API, and no adapter for any other model family. Researchers without Apple Silicon and Qwen3.6-27B (4-bit) specifically cannot run it at all.
Bottom line: Pick this for hands-on mechanistic interpretability work on Qwen3.6-27B on Apple Silicon — abandon it the moment your target model is anything else or your hardware is not MLX-capable.
Community Performance Report Card
No community ratings yet. Be the first to rate this tool!
Community Benchmarks Community
Sign in to submit a benchmarkNo community benchmarks yet. Be the first to share a real-world data point.
Pros
Sign in to edit- Layer-by-layer Jacobian lens visualization surfaces what the model's intermediate representations are predicting at every token position, so you can catch the gap between a model's output and its internal trajectory — the gap that final-logit inspection alone cannot show.
- Fully local and Apache-2.0 licensed, which means your prompts, model weights, and interpretability results never leave your machine — critical when the inputs are sensitive or proprietary.
- A browser-hosted read-only demo at jlens.wezzard.com lets you evaluate the visualization interface before committing to local installation, so you do not spend setup time on a workflow that does not match your needs.
- Custom Jacobian lens fitting is supported, so researchers can experiment with their own linear correction layers rather than being locked to a single pre-fitted probe.
- Self-hosted and dependency-managed via uv with a full test suite included, which means reproducible local environments without fighting conflicting package versions.
Cons
Sign in to edit- The tool runs only on Qwen3.6-27B (4-bit) via MLX on Apple Silicon — there is no documented adapter path for any other model architecture or weight format. Teams whose target model is GPT-based, Llama-family, Mistral, or anything else have no migration path and must look at tools like TransformerLens or NNsight instead.
- There is no API surface. Every interaction is through the local web UI. Teams that want to script interpretability checks into a CI pipeline, run batch prompt comparisons programmatically, or integrate lens outputs into a larger evaluation harness have to build that scaffolding themselves against the Python internals, with no documented public interface to depend on.
- Hardware without Apple Silicon cannot run this at all. CUDA-based workstations, Linux servers, and cloud GPU instances are out of scope. Research teams without M-series Macs are blocked at the dependency layer before they reach any model-specific constraints.
Community Reviews
Sign in to write a reviewNo reviews yet. Be the first to share your experience.
About
- Platforms
- Apple Silicon Mac
- API Available
- No
- Self-Hosted
- Yes
- Last Updated
- 2026-07-14T14:20:47.068Z
Best For
Who it's for
- Researchers studying internal model representations
- Developers debugging local MLX-based Qwen models
- Users wanting layer-wise token visualization on Apple Silicon
What it does well
- Inspecting latent concepts during model generation
- Debugging refusal or hidden reasoning in LLMs
- Comparing per-layer token predictions visually
- Fitting and testing custom Jacobian lenses
Integrations
Discussion Community
Sign in to commentNo discussion yet. Sign in to start the conversation.
Compare J-lens Qwen
Spotted incorrect or missing data? Join our community of contributors.
Sign Up to ContributeCommunity Notes & Tips Community
Sign in to contributeBe the first to contribute. General notes, observations, gotchas, and tips from people who use this tool day-to-day.
Frequently Asked Questions
- Is J-lens Qwen free?
- Yes — J-lens Qwen is fully free to use. There is no paid tier.
- Is J-lens Qwen open source?
- Yes. J-lens Qwen is open source.
- Can I self-host J-lens Qwen?
- Yes. J-lens Qwen supports self-hosting on your own infrastructure.
- What platforms does J-lens Qwen support?
- J-lens Qwen is available on: Apple Silicon Mac.
Hours Saved & ROI Stories Community
Sign in to contributeBe the first to contribute. Concrete time/cost savings, with context. e.g. "Cut my code review backlog from 4h to 45m per week."
Best J-lens Qwen alternatives →
Curated lists that include this category
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.