Neshter: a fully local tool to visualize, measure and prune attention heads in small LLMs
Bayqus · reddit · 2026-11-22
A developer released Neshter (Apache-2.0), a 100% local tool for inspecting attention heads in small language models—no cloud API, no account, no telemetry.
Key features:
- Load a model, see a heatmap of each head's contribution to the next token, pick heads, and measure the effect of switching them off against a random-head baseline from the same layers; save the result as a new model file.
- One-click example: on Qwen2.5-0.5B-Instruct, disabling the two strongest heads for 'The capital of France is' drops ' Paris' probability from 30% to 25%. The point is the random baseline—a head that looks important on the map isn't always important when actually cut.
- A 'find a behavior' mode ranks heads carrying a behavior from positive/negative examples and measures each; 3D views of the heatmap and information flow included.
Limits: designed for models up to 8B (author mostly uses 0.1-3B), requires eager attention, falls back to CPU, tested on Windows and Linux, Android build from source only, v0.2 one-person project.
More from Research
- 10 months after a Reddit question on AI solving Millennium Prize Problems, claims abound — haider1 · 2026-11-22
- Fable 5.2 rewrites model weights so an E2E driving bot learns left-side driving — 4310sy · 2026-11-22
- IceCube citizen science challenge draws 4,000+ classifications to help train AI — CatAstro_Piyush · 2026-11-22
- Michael Levin teases Morphoceuticals, his bioelectricity regeneration startup — drmichaellevin · 2026-11-22
- Reddit manifesto argues cybernetics shows humanity never escaped nature — Cyborgized · 2026-11-22
- Pedro Domingos: Masullo lost $1M by having his proof checked by a human instead of an LLM — pmddomingos · 2026-11-22