---
title: Divinci AI
emoji: 🧠
colorFrom: green
colorTo: yellow
sdk: static
pinned: false
short_description: Feature-level interpretability for open transformers
---

# Divinci AI

Feature-level interpretability artifacts for open transformers — built openly, validated empirically.

A **vindex** is a transformer's weights decompiled into a queryable feature database. It exposes the entity associations, circuit structure, and knowledge-editing surfaces that live inside a model's FFN layers — without requiring GPU inference for most operations.

Think of it as the model's index: the thing you search before you run it.

---

## Interactive viewer

[![LarQL Vindex Viewer — interactive 3D + 2D circuit visualization](https://huggingface.co/spaces/Divinci-AI/vindex-viewer/resolve/main/vindex-hero-bg.gif)](https://huggingface.co/spaces/Divinci-AI/vindex-viewer)

**[→ Open the interactive viewer](https://huggingface.co/spaces/Divinci-AI/vindex-viewer)**

Pick any of 9 models from the dropdown. Toggle between the 3D cylinder spiral and a flat 2D circuit/network view. Hit **⇌ Compare** to render the current model alongside Bonsai 1-bit, side-by-side — the contrast between fp16 structure (organized rings) and 1-bit dissolution (scattered cloud) is the most direct picture of what 1-bit training does to a transformer's internal organization that we know how to render. Search for entity features (`?q=paris&model=gemma-4-e2b`) to see real probe-derived activations light up across the layer stack — backed by a 5000-token offline-built search index.

---

## Published vindexes

Cross-family evidence in hand: **Gemma**, **Qwen3**, **Mistral**, **Llama**, **OpenAI MoE**, **Moonshot MoE**, **DeepSeek-V4 MoE**, plus two 1-bit controls.

<table>
<tbody>
<tr><td><strong>MODEL</strong></td><td><strong>ARCHITECTURE</strong></td><td><strong>PARAMS</strong></td><td><strong>VINDEX</strong></td><td><strong>C4 / var@64</strong></td><td><strong>STATUS</strong></td><td><strong>NOTES</strong></td></tr>
<tr><td><strong>Gemma 4 E2B-it</strong></td><td>Dense (Gemma 4)</td><td>2B</td><td><a href="https://huggingface.co/Divinci-AI/gemma-4-e2b-vindex">gemma-4-e2b-vindex</a></td><td><strong>0.0407 ± 0.0004</strong> ✓</td><td>Complete</td><td>3-seed validated; headline universal-constant model</td></tr>
<tr><td>Qwen3-0.6B</td><td>Dense (Qwen 3)</td><td>0.6B</td><td><a href="https://huggingface.co/Divinci-AI/qwen3-0.6b-vindex">qwen3-0.6b-vindex</a></td><td>0.411</td><td>Complete</td><td>Smallest published; Qwen3 family-elevated C4</td></tr>
<tr><td>Qwen3-8B bf16</td><td>Dense (Qwen 3)</td><td>8B</td><td><a href="https://huggingface.co/Divinci-AI/qwen3-8b-vindex">qwen3-8b-vindex</a></td><td>0.804</td><td>Complete</td><td>Architecture control for Bonsai</td></tr>
<tr><td>Qwen3.6-35B-A3B</td><td>MoE (Qwen 3.6)</td><td>35B / 3B active</td><td><a href="https://huggingface.co/Divinci-AI/qwen3.6-35b-a3b-vindex">qwen3.6-35b-a3b-vindex</a></td><td>—</td><td>Complete</td><td>256 experts, 40 layers</td></tr>
<tr><td>Ministral-3B</td><td>Dense (Mistral 3)</td><td>3B</td><td><a href="https://huggingface.co/Divinci-AI/ministral-3b-vindex">ministral-3b-vindex</a></td><td>0.265</td><td>Complete</td><td>Post-quant fp8 → bf16; non-dissolved spectrum</td></tr>
<tr><td>Llama 3.1-8B</td><td>Dense (Llama 3.1)</td><td>8B</td><td><a href="https://huggingface.co/Divinci-AI/llama-3.1-8b-vindex">llama-3.1-8b-vindex</a></td><td><strong>0.012</strong> ✓</td><td>Complete</td><td>Llama family signature</td></tr>
<tr><td>MedGemma 1.5-4B</td><td>Dense (Gemma multimodal)</td><td>4B</td><td><a href="https://huggingface.co/Divinci-AI/medgemma-1.5-4b-vindex">medgemma-1.5-4b-vindex</a></td><td><strong>1.898 ⚠</strong></td><td>Complete</td><td>45× cohort anomaly — under investigation</td></tr>
<tr><td>GPT-OSS 120B</td><td>MoE (OpenAI)</td><td>120B</td><td><a href="https://huggingface.co/Divinci-AI/gpt-oss-120b-vindex">gpt-oss-120b-vindex</a></td><td>—</td><td>Complete</td><td>S[0] grows 117× with depth (L0=111 → final=13,056)</td></tr>
<tr><td><strong>Kimi-K2-Instruct</strong></td><td>MoE fp8-native (DeepSeek-V3 style)</td><td>1T / 32B active</td><td><a href="https://huggingface.co/Divinci-AI/kimi-k2-instruct-vindex">kimi-k2-instruct-vindex</a></td><td><strong>0.0938</strong> (MoE median) ‡</td><td>Complete</td><td>60 MoE layers; 42.28 GB gate_proj binary; broader L52–L60 secondary rise than initial dome SVD suggested</td></tr>
<tr><td><strong>DeepSeek-V4-Flash</strong></td><td>MoE MXFP4 (DeepSeek-V4)</td><td>43L / 256 experts / 6 active</td><td><em>publishing soon</em></td><td><strong>—</strong></td><td><strong>Phase 1B running</strong></td><td>43-layer all-MoE; first-peak L17 + double-bend profile (distinct from Kimi’s smooth dome); MXFP4 unpacker added to builder</td></tr>
<tr><td><strong>DeepSeek-V4-Pro</strong></td><td>MoE MXFP4 (DeepSeek-V4)</td><td>61L / 384 experts / 6 active</td><td><em>queued</em></td><td>—</td><td>Queued</td><td>Same scale as Kimi-K2 (60–61 layers × 384 experts × 7168 hidden); MXFP4 expert weights</td></tr>
<tr><td><strong>Bonsai 8B</strong></td><td>1-bit (Qwen 3 base, post-quantized)</td><td>8B</td><td><em>vindex pending publish</em></td><td>0.093 (var@64)</td><td>Phase 1 complete</td><td><strong>C5 = 1</strong> (circuit dissolved); n=1 of 1-bit dissolution</td></tr>
<tr><td><strong>BitNet b1.58-2B-4T</strong></td><td>1-bit (Microsoft, native)</td><td>2B</td><td><em>vindex pending publish</em></td><td>0.111 (var@64)</td><td>Phase 1 complete</td><td>n=2 dissolution confirmation; native 1-bit training</td></tr>
</tbody>
</table>

‡*Kimi-K2 final: 60 MoE layers (L01–L60), gate_proj SVD, median var@64=0.0938 (range 0.083–0.108). Phase 1 + Phase 1B + Phase 2 all complete 2026-04-24; 42.28 GB binary published. DeepSeek-V4 series builds with MXFP4 unpacker (V4-Flash 1B in progress 2026-04-25, V4-Pro queued). Card updates in-place as phases land.*

---

## What's a vindex?

Standard model weights tell you *what* a model computes. A vindex tells you *where* it stores specific knowledge and *which features* need to change for a targeted edit.

Concretely: given a query like `"Paris → capital"`, a vindex walk returns the layers, feature directions, and token associations that encode that fact. A patch operation writes a rank-1 ΔW that suppresses or overwrites that association — compiled back to standard HuggingFace safetensors for inference.

LarQL (the toolchain that builds vindexes) is open-source: [github.com/chrishayuk/larql](https://github.com/chrishayuk/larql) | [github.com/Divinci-AI/larql](https://github.com/Divinci-AI/larql).

---

## Research

### Paper 1 — *Architectural Invariants of Transformer Computation*
*arXiv preprint forthcoming*

Five properties measured across every model in this collection. **Three hold within ±15% coefficient of variation** across architectures, organizations, and scales. **One collapses under 1-bit quantization** — replicated across two independent 1-bit models from two organizations (n = 2). **One scales monotonically with model size**.

The headline universal constant — layer temperature C4 — is reproducible at the **1% precision level**: a three-seed run on Gemma 4 E2B gives `C4 = 0.0407 ± 0.0004`, with circuit-stage count perfectly stable (`C5 = 4 ± 0`) across all seeds.

### Paper 2 — *Constellation Edits*
*draft, arXiv after 3-seed runs + α-sweep appendix*

Mechanistic knowledge editing in transformer feature space. Includes a negative result: why activation-space edits fail in 1-bit models, and what weight-space geometry reveals about why.

### Companion blog series — *The Interpretability Diaries*

- [Part I — The Architecture Every Language Model Converges To](https://divinci.ai/blog/architecture-every-llm-converges-to/) — five universal constants, what holds and what doesn't
- [Part II — Deleting Paris from a Language Model](https://divinci.ai/blog/deleting-paris-from-a-language-model/) — Gate-3 surgical knowledge edit with a receipt; rank-1 ΔW that suppresses one fact at +0.02% perplexity
- [Part III — When the Circuit Dissolves](https://divinci.ai/blog/when-the-circuit-dissolves/) — three dissolution datapoints (BitNet, Bonsai, Kimi-K2): var@64 ≈ 0.09–0.10 for 1-bit + fp8-native vs ~0.85 for fp16/post-quant. Training precision, not storage precision, predicts spectral structure.

Working notebooks: [github.com/Divinci-AI/server/tree/preview/notebooks](https://github.com/Divinci-AI/server/tree/preview/notebooks)

---

## Working in public

Every measurement in our papers traces back to a notebook and a commit. Negative results ship alongside positive ones — the MLP compensation mechanism that defeats knowledge editing in 1-bit models is in the notebooks, not buried in a supplement.

If you replicate a result and find a discrepancy, open an issue on the LarQL repo.

---

*Vindexes on this org are free for academic and research use (CC-BY-NC 4.0). Commercial licensing: mike@divinci.ai*