The model is the database: LARQL runs and interrogates LLMs as artifacts

What it does

LARQL is a Rust model runtime and the reference implementation of VINDEX3, a format that captures a model’s logical structure, physical encodings, and executable semantics in one artifact. You encode a Hugging Face checkpoint into a .vindex3 file, then run it, record its computation, inspect its internals, and test claims about its behavior. It also serves via an OpenAI-compatible API surface, per its topics.

The interesting bit

The thesis is that the model itself is the thing you query — no separate database of extracted facts. Decodes leave execution records with site identity and provenance, which the Observatory tool imports for coordinated inspection and replay. That’s a genuinely different shape from the usual “dump activations and hope” interpretability workflow.

Key highlights

• Encode, run, observe, intervene, and query a model as a single versioned artifact
• Execution records with provenance feed head/source attribution and intervention research
• REPRESENT compiles alternate encodings (e.g. NVFP4) with measurement contracts comparing candidate against reference
• Legacy VINDEX2 workflows survive: LQL queries, patch overlays, compilation, Python bindings
• CPU and Metal backends; other encodings and providers load via --plugin shared libraries

Caveats

• Admission and backend support are explicit gates — not every checkpoint will execute
• Sharded/partial execution and multimodal embedding handoff are planned and explicitly refused until contracts exist
• The README is candid that fixture passes aren’t model-wide fidelity or speed claims; historical benchmarks stay in research records

Verdict

Worth a look if you care about mechanistic interpretability or model editing and want a runtime built for evidence rather than vibes. If you just need a fast inference server, this is a research system first — the docs themselves distinguish supported interfaces from active research.