jlens-gguf brings Anthropic-style Jacobian lens to llama.cpp GGUF models
Steer and visualize internal states of 160GB GGUF models with 20GB extra RAM
A new open-source tool called jlens-gguf brings Anthropic's Jacobian lens technique to the GGUF ecosystem and llama.cpp. Built by developer Igor Barshteyn with heavy AI assistance from Fable 5, the project provides a native GGUF server (closely synced to llama.cpp) that lets users not only observe but actively steer model internals via Jacobian-space swapping, ablitteration, and steering. This fills a gap where prior Jacobian-lens projects only supported HuggingFace and PyTorch models.
The tool's memory requirements scale at about 1/8 of model size — so a 160GB model like Qwen3.5-397B UD-Q3_K_XL needs roughly 20GB additional RAM for the lens. It supports both dense and mixture-of-experts (MoE) GGUFs. While it can observe already-running llama-server models in read-only mode, full steering requires jlens-gguf's own server. The project is available on GitHub and references Anthropic's global workspace paper and code as inspiration.
- Provides both observation and active steering of GGUF model internals via Jacobian-space manipulation
- Memory overhead is ~12.5% of model size (e.g., 20GB extra for a 160GB Qwen3.5-397B)
- Works with dense and MoE GGUFs; includes a native GGUF server synced to llama.cpp
Why It Matters
First practical tool to bring Anthropic-level mechanistic interpretability to local GGUF models on consumer hardware.