# Danger Labs: dm-jepa

> A 308M non-autoregressive choice model: a ModernBERT-base encoder reads up to 16K tokens of state, a small latent predictor scores each option, and a softmax gives a probability per option in one forward pass. Needs the custom code in its repo; weak on held-out tasks.

- Page: https://systemonemodels.ai/danger-labs/dm-jepa
- API: https://api.systemonemodels.ai/v1/models/danger-labs/dm-jepa
- Download: `pip install systemonemodels && systemone pull danger-labs/dm-jepa`

## Facts

| | |
|---|---|
| Maker | Danger Labs (https://systemonemodels.ai/danger-labs) |
| Decides | choice |
| Architecture | dm-jepa |
| Base model | answerdotai/modernbert-base |
| Parameters | 308M |
| Context | 16K tokens |
| Licence | mit |
| Availability | Open weights |
| Released | 2026-10-04 |
| Latest version | 1.1.0 |

## Reported evaluation

Suite: Decision Index 0.2.1 full panel (150,759 requests), maker's self-run. Numbers are the publisher's own.

- Median latency: 34.3 ms

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# DM-JEPA 1.1

A 308M non-autoregressive choice model: a ModernBERT-base encoder reads up to 16K tokens of state, a small latent predictor scores each option, and a softmax gives a probability per option in one forward pass. Needs the custom code in its repo; weak on held-out tasks.

Danger Labs calls it a JEPA-style "System 1" decision engine. A ModernBERT-base encoder, with RoPE stretched to 16,384 tokens of state and 512 per option, embeds the state and each option; a four-step recurrent predictor scores each option, and a temperature-scaled cosine softmax returns a probability per option. Choice questions only. The card says "calibrated" but publishes no calibration error. It will not load through transformers AutoModel: it needs modeling_dm_jepa.py and the djepa/ package in the repo, plus the ModernBERT-base tokenizer. The maker reports its own run of the Decision Index 0.2.1 protocol: 23.16 balanced skill and 41.92 raw over 150,759 requests. The same card claims 100% on GSM8K and on an "OpenJev High-Trust" suite, which sit oddly with that overall score. The third-party Decision Index 0.3 board scores DM-JEPA 1.1 at 4.05 (rank 105), and 2.84 on new domains.

## What it decides

- **choice** — picks one option from a set

## At a glance

| | |
|---|---|
| Parameters | 308M |
| Base model | `answerdotai/ModernBERT-base` |
| Maker | Danger Labs |
| Released | 2026-10-04 |
| License | mit |
| Reported latency | 34.3 ms median device-synchronised forward pass on one NVIDIA GB10 |

## Get the weights

```bash
pip install systemonemodels
systemone pull danger-labs/dm-jepa
```

The files are served from the maker's Hugging Face repository, [`DangerLabs/DM-JEPA`](https://huggingface.co/DangerLabs/DM-JEPA), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/DangerLabs/DM-JEPA)
- [Decision Index results (maker's run)](https://huggingface.co/datasets/DangerLabs/decision-index-results)

---

*This page was opened by System One for Danger Labs, who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models