NOTICE

collabosm - NOTICE

This repository contains NO model weights. Everything in scripts/ and docs/ is ours (MIT).

Model
  Qwen3.8-Flash-Next exl3 quantizations by turboderp
    https://huggingface.co/turboderp/Qwen3.8-Flash-Next-exl3
    revision used here: 4.05bpw_h6_ng6
    sha: 55a732e0c4c3d4614bc42b68493bb930d9b02c0a
  The weights carry their own licence, published in the repository above. Read it before you
  redistribute anything derived from them, and note that serving a model to others is a separate
  question from downloading it. We link and pin; we do not mirror.

Runtime
  ExLlamaV3 (MIT) - https://github.com/turboderp-org/exllamav3
    pinned: v1.5.1, prebuilt wheel referenced in manifest.json

Tooling we depend on at run time, but do not vendor
  google-colab-cli (Apache-2.0) - https://github.com/googlecolab/colab-cli
  cloudflared (Apache-2.0)       - quick tunnels only, for reaching a localhost port
  huggingface_hub (Apache-2.0)

Measurements in docs/MEASURED.md were produced by this kit on a single Colab A100-80GB High-RAM.
Numbers from other people's hardware are attributed where used.