MetaEncoder-30B
===============

This work is a derivative of Muse Glimmer-30B, (c) Meta Platforms, Inc., obtained from
https://huggingface.co/meta-models/Muse-Glimmer-30B and licensed under the Apache License,
Version 2.0. A copy of that license is distributed with this work as LICENSE.

Trademarks
----------

The Apache License, Version 2.0 grants no trademark rights (Section 6). "Meta", "Muse", and
"Muse Glimmer" are trademarks of Meta Platforms, Inc. The name "MetaEncoder" is used here only
to identify this derivative work and does not indicate that it is an official Meta product or
that Meta Platforms, Inc. endorses it.

Statement of changes (Apache License 2.0, Section 4(b))
-------------------------------------------------------

The following files differ from the corresponding files in the base model:

  model-00001-of-00002.safetensors
  model-00002-of-00002.safetensors
  model.safetensors.index.json
      The weights of the language tower have been modified by contrastive fine-tuning for
      embedding extraction. A LoRA adapter (rank 128, alpha 256) trained on the
      self_attn.{q,k,v,o,gate}_proj and mlp.{gate,up,down}_proj projections of all 52 language
      layers was merged into the base weights. The perception encoder, token embeddings,
      lm_head, and all normalisation layers are unmodified and remain bit-identical to the
      base model.

  config.json
      Regenerated on save. Differs from the base only in recorded token ids, dtype fields,
      use_cache, and the transformers version string.

  processor_config.json
      max_image_tokens reduced from 4096 to 1280 to match the budget the model was trained
      under.

  README.md
      Replaced with a model card for this derivative.

The following files are new in this work and are not part of the base model:

  modeling_metaencoder.py
      Inference wrapper.

  USAGE.md
      Usage guide for this derivative.

  requirements.txt
  .gitattributes
      Packaging metadata for this derivative.

  NOTICE
      This file.

All other files (tokenizer.json, tokenizer_config.json, chat_template.jinja,
generation_config.json, LICENSE, USAGE_POLICY.md) are unmodified copies from the base model.

MLX conversion: edgefloor, 2026-10-09. Language weights quantized to affine 4-bit/group-64; vision weights preserved. Added MLX inference and conversion scripts.