Models / aurora-135m-v1

aurora-135m-v1 stable

Public release. Instruction-tuned 135M, ChatML format. This is the build we benchmark against.

6 files main 272.6 MB
chat_template.jinja Sep 20 368 Bytes
config.json Sep 20 905 Bytes
generation_config.json Sep 20 132 Bytes
model.safetensors Sep 20 269.1 MB
tokenizer.json Sep 20 3.5 MB
tokenizer_config.json Sep 20 453 Bytes
CLI curl -LO http://nightshift.codefest.ctfd.live/models/aurora-135m-v1/resolve/model.safetensors

Model card

  • Base: SmolLM2-135M-Instruct (HuggingFaceTB)
  • Released: 2026-08-30 by registry
  • ChatML template, fp16 safetensors, vocab 49152
  • Status: stable, externally distributed
  • Playground: no

Release notes

aurora-135m-v1: public release

2026-08-30 ยท registry

aurora-135m-v1 is the stable public build of our in-house assistant line: a straight export of SmolLM2-135M-Instruct with the ChatML template, fp16 weights in a single safetensors file, and a 49152-token vocabulary.

v1 is our benchmark baseline. Every nightly calibration pass is scored against it before it lands on the registry. The tag is frozen: do not fine-tune on it, and do not expect it to change.

v1 does not run in the playground. It is here to download and to diff against.