Outlier-40B / README.md
ur-dad-matt's picture
docs(hf): standardize Outlier-40B discoverability metadata + model card
7e02479 verified
metadata
language:
  - en
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: Qwen/Qwen2.5-14B-Instruct
base_model_relation: adapter
tags:
  - text-generation
  - mixture-of-experts
  - moe
  - ternary
  - qwen2
  - qwen2.5
  - outlier
  - legacy
  - research
  - superseded

Outlier-40B — V1 (legacy)

Superseded. This is the V1 release of Outlier-40B, preserved for reproducibility. Use Outlier-40B-V3.2 for current work.

Why this repo still exists

  • Reproducibility of early experiments and papers that cite this version.
  • Historical interest — this was part of the Day-17 2026 Outlier launch.

What changed in later versions

V3.2 and later apply the alpha-fix scalar correction (+1.61pp at 70B per registry E05) and incorporate the expert-paging three-tier memory architecture. Legacy releases do not.

V1 release. Use V3.2 for current work.

See also

Citation

@misc{outlier2026,
  author       = {Kerr, Matt},
  title        = {Outlier: Ternary Mixture-of-Experts for Consumer Hardware},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Outlier-Ai}}
}

Links