---
license: apache-2.0
language:
  - en
  - zh
library_name: pytorch
tags:
  - spiking-neural-networks
  - language-model
  - sparse-activation
  - inference-only
  - pytorch
---

# SymbolicLight V1

This repository contains the inference-only model artifacts for SymbolicLight
V1, described in [arXiv:2605.21333](https://arxiv.org/abs/2605.21333).

The compatible model definitions, tokenizer wrappers, and minimal generation
entry points are maintained in the
[public code repository](https://github.com/SymbolicLight-AGI/SymbolicLight-V1).

## Public scope

This repository contains exactly one released checkpoint for each model scale,
plus its matching tokenizer:

| Model | Checkpoint | Bytes | SHA-256 |
| --- | --- | ---: | --- |
| 0.8B | `0.8B/weights/pytorch/latest-inference.pt` | 3,500,562,525 | `5edc3b6a6d3fb9cfaf713775133361b3ff907abf293a22ef2226785f9fbc0540` |
| 194M | `194M/weights/pytorch/symboliclight-v1-194m-inference.pt` | 776,134,571 | `871dfe12d082d896dd9f5e55186f9fb7397ba7ed2b01953f8234b3e575fce521` |

The checkpoint containers retain only model tensors and configuration fields
required to construct the inference graph. They do not contain optimizer or
gradient-scaler state, training steps, token or data counters, best-loss
values, data-loader state, training datasets, local paths, or training-result
metadata.

Training, ablation, baseline, and intermediate checkpoints are not distributed.
The public release supports checkpoint loading and inference reproduction; it
does not claim end-to-end reproduction of training or paper metrics.

The 0.8B and 194M tokenizers are different and must not be interchanged. See
the model-specific READMEs and `SHA256SUMS` before loading an artifact.

## License

Unless a file states otherwise, the released SymbolicLight weights, tokenizer
assets, documentation, and release metadata are licensed under Apache License
2.0. Training and validation corpora are not included or licensed here. See
`WEIGHTS_LICENSE.md` for the release boundary.