NNterp 2.0 Launches with Unified Interpretability API and Tooling Ecosystem
On November 19, davidbau's team launched NNterp 2.0 along with an accompanying tooling ecosystem, aiming to give mechanistic interpretability research a unified infrastructure where "experiment code is written once and runs across many models." At its core is the NNterp 2.0 API: it provides a unified syntax for accessing model components such as decoder heads, residual streams, and attention, abstracting away the quirks of various transformer architectures so that the same piece of NNsight experiment code runs on virtually all LLMs.
Confirmed
- NNterp 2.0 supports architectures covering modern variants including linear attention, SSMs, MoE, and vision encoders (m1, m4).
- The NNterp Encyclopedia is live, serving as a visual quick-reference for modern transformer architecture variants; combined with NDIF, it enables research on hundreds of open-weight models, addressing the problem of hard-to-remember architectural differences (m2).
- NNsider Workbench has been released: a visual tool built on NNterp and NDIF that requires no coding—complex intervention experiments on any HuggingFace model can be assembled via point-and-click (m3).
- NNsight supports remote execution: setting remote = True in a model tracing call loads the model onto NDIF servers, capable of running 70B models, with NDIF providing free NSF-funded compute (m5).
Why it matters
- For mechanistic interpretability researchers, NNterp 2.0 substantially lowers the engineering cost of running cross-model experiments (per m4 author gsarti).
- Bau noted that as interpretability methods like probing become routine tools in production-grade AI, researchers can carry out this work directly on NNsight/NDIF, and free compute lowers the barrier to large-scale model research (m5).
- The Encyclopedia and Workbench respectively lower the barriers to "understanding architectural differences" and "getting started with intervention experiments," completing a full tooling ecosystem alongside the API.
2026-11-19 ~ 2026-11-19 · 5 related posts
Primary sources
- NNterp 2.0 API released: one unified interface for interpreting attention, SSM, MoE and vision models — davidbau ·
- Workbench: browser-based LLM intervention experiments on HuggingFace models, no code needed — davidbau ·
- NNsight interpretability tool runs 70B models remotely on free NSF-backed NDIF GPUs — davidbau ·
- NNterp 2.0 launches: run the same NNsight interpretability experiment across ~all LLMs — gsarti_ · 2026-11-19
- NNterp Encyclopedia: a visual guide to every modern transformer architecture variant — davidbau · 2026-11-19
- [source] NNterp 2.0 API released: one unified interface for interpreting attention, SSM, MoE and vision models — davidbau · 2026-11-19
- [source] Workbench: browser-based LLM intervention experiments on HuggingFace models, no code needed — davidbau · 2026-11-19
- [source] NNsight interpretability tool runs 70B models remotely on free NSF-backed NDIF GPUs — davidbau · 2026-11-19