NNterp 2.0 API released: one unified interface for interpreting attention, SSM, MoE and vision models
davidbau · x · 2026-11-19
davidbau's team released the NNterp 2.0 API, a unified idiom for accessing decoder heads, residual streams, attention and other components, so an experiment written once runs across many models. It supports standard attention, MLPs, mixture-of-experts, linear attention, state-space mixers and vision encoders; the quoted example shows the same code working across 100 transformer model families by swapping the repo. Combined with NDIF, researchers can study the internals of hundreds of open-weight models. The release is accompanied by the NNterp Encyclopedia, a visual guide covering block diagrams, quirks and interp-specific notes for each architecture family.
Related event: NNterp 2.0 Launches with Unified Interpretability API and Tooling Ecosystem(5 posts)→
More from Research
- tessera-small: a diffusion-LLM-based pixel-art text-to-image model — thejagstudio · 2026-11-19
- WorldRuler: rule-graph world models that honor user-defined game rules — AlayaLab · 2026-11-19
- NexSight: verifiable long-video tasks show top models score only 54.5% — nex-agi · 2026-11-19
- PReMind: benchmark + RL framework for hazard-aware personalized VLM assistants — SeoulNatlUniv · 2026-11-19
- Late Entry in looped LLMs delivers up to 12.5x lossy and 8.4x exact speedups — NanyangTechnologicalUniversity · 2026-11-19
- Optimization × AI Forum on Nov 20: two Peking University talks on AI-driven math discovery — aaron_defazio · 2026-11-19