Skip to content

ARC_NEURON_MODEL_FAMILY.md

Latest commit

 

History

History
39 lines (32 loc) · 1.52 KB

ARC_NEURON_MODEL_FAMILY.md

File metadata and controls

39 lines (32 loc) · 1.52 KB
Raw

ARC-Neuron Model Family

ARC-Neuron is the model family built around ARC Cognition Core.

Current baseline

  • ARC-Neuron Tiny 0.05M v0.1 (F32, GGUF)
  • Purpose: end-to-end proof that the ARC stack can train, export, validate, load, and run a local GGUF artifact.
  • Status: working local baseline, not a flagship capability model.

Family ladder

  • ARC-Neuron Tiny — lab proof, container validation, runtime smoke, regression floor
  • ARC-Neuron Base — first usable small instruction model
  • ARC-Neuron Core — main local general model
  • ARC-Neuron Command — command routing, planning, tool discipline
  • ARC-Neuron Language — language and transliteration specialization
  • ARC-Neuron Native — deployment build packaged for local native runtime

What Tiny proves

  1. Repo-local training can emit a real .gguf file.
  2. The .gguf can be validated without external services.
  3. The runner can load the artifact and generate text locally.
  4. ARC Cognition Core can act as the control plane around the artifact.

What Tiny does not claim

  • competitive frontier capability
  • llama.cpp compatibility for this toy architecture
  • a finished production flagship

Release doctrine

Every ARC-Neuron release should ship with:

  • a model card
  • a GGUF artifact
  • a validation receipt
  • a benchmark report
  • a provenance manifest
  • a runtime launch path

Next hard step

Promote from ARC-Neuron Tiny to ARC-Neuron Base by plugging a stronger upstream checkpoint into the production build path already present in this repo.