---
license: mit
library_name: onnxruntime
tags:
  - onnx
  - webgpu
  - continuous-thought
  - coconut
---

# Neuralese COCONUT browser graph

This repository hosts the compact ONNX graph used by the **Neuralese realtime
latent instrument**. It is derived from the MIT-labelled
[`ModalityDance/latent-tts-coconut`](https://huggingface.co/ModalityDance/latent-tts-coconut)
checkpoint at revision `89501ce8cc7cefd020b6e30c475a6772494de0a4`.

The graph implements strict COCONUT recurrence: a decoder final hidden state can
be supplied directly as the next input embedding before returning to token mode.
It is not an ordinary token model with post-hoc activation sonification.

## Artifact

`coconut-decoder-fp16-storage.onnx`

- 124M-parameter GPT-2 architecture
- 12 layers, 12 heads, hidden size 768
- FP16 weight storage with FP32 inputs, arithmetic, outputs, and KV cache
- SHA-256: `3bc35680260edb1c66276bbd65ae347535ec89ccb80914b469c1d8cf55f400bd`
- Source graph: 622 MiB FP32
- Compact graph: 311 MiB

Native FP16 arithmetic was tested but rejected because its recurrent trajectory
fell below the project's browser fidelity gate. The published graph halves the
download while preserving FP32 computation.

## Measured browser parity

Against the publisher-faithful PyTorch trajectory for the fixed six-state probe:

- identical generated tokens;
- minimum latent cosine: `0.99999919`;
- maximum absolute latent error: `1.742e-2`;
- approximately 29–30 ms per model call after warm-up on the development M4 Pro.

The checkpoint's answer to the probe is incorrect (`### 6` for `2 + 2`). This is
a checkpoint capability result, not a conversion discrepancy. The artifact is
published for artistic research, browser inference, and reproducibility—not as
a reliable general reasoning model.

## Attribution

Model weights originate from `ModalityDance/latent-tts-coconut`. The recurrent
continuous-thought method is based on COCONUT, *Training Large Language Models
to Reason in a Continuous Latent Space* (Meta FAIR, 2024).