xLLM Part II model artifacts
Towards Looped Models Done Right, Part II: Rethinking at Fixed Points

The model weights are distributed under the Apache License, Version 2.0.
A copy is provided in LICENSE. The xLLM code has its own license in the
repository root; third-party code retains its original notices.

Tokenizer attribution

The tokenizer is the one the models were trained with: a BPE tokenizer with a
64,000-token vocabulary that the authors' team trained on a Jais-style data
mix, which the paper calls Jais64k after Jais (Sengupta et al., 2023), with
added chat and tool special tokens. It is distributed with the weights under
the Apache License, Version 2.0, in tokenizer/: tokenizer.json,
tokenizer_config.json and special_tokens_map.json are copied unchanged. A
distilled student has no tokenizer/ and uses its teacher's.

Artifact preparation:
- The trained FP32 tensors are rounded to BF16 and stored in Safetensors
  shards, keyed as the xLLM model's state dict.
- config.json holds the release model config; for models with a tokenizer, its
  tokenizer section points to tokenizer/, and for distilled students its student
  section names the teacher artifact and its manifest SHA-256.