# Blackdrome AI Labs: noma

> Open decision model from Blackdrome AI Labs: the first 18 of 32 layers of Qwen3.5-4B-Base plus trained option heads. For choice, noul and score questions it returns a probability per option, an abstain signal and ensemble uncertainty, in one pass. Serves /v1/systemone.

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

## Facts

| | |
|---|---|
| Maker | Blackdrome AI Labs (https://systemonemodels.ai/blackdrome-ai-labs) |
| Decides | choice, score, noul, classify, route |
| Architecture | noma |
| Base model | qwen/qwen3.5-4b-base |
| Parameters | 2.7B |
| Context | 4K tokens |
| Licence | mpl-2.0 |
| Availability | Open weights |
| Released | 2026-10-01 |
| Latest version | 2026.09 |

## Reported evaluation

Suite: Blackdrome sealed set (386 human-reviewed decisions, 12 families, never used in training; the maker's own). Numbers are the publisher's own.

- Decision accuracy: 82.6%
- Calibration error (ECE): 0.042
- Median latency: 16 ms

## Model card

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

# Noma

Open decision model from Blackdrome AI Labs: the first 18 of 32 layers of Qwen3.5-4B-Base plus trained option heads. For choice, noul and score questions it returns a probability per option, an abstain signal and ensemble uncertainty, in one pass. Serves /v1/systemone.

A single safetensors file holds 18 layers with a merged rank-16 LoRA and four bootstrap listwise option scorers that read each option's marker token; a deterministic fact channel supplies dates and quantities. English only, states up to 4,096 tokens, NVIDIA GPUs only. Training used about 34k decisions from permissive public data, code-generated items and blind-labelled items. On its own sealed set (386 human-reviewed decisions never used in training) Blackdrome reports 82.6% with ECE 0.042. On JevBench's 231 public items (a third-party benchmark, the maker's own run, not yet on the board) it reports 76.2% overall but 51.4% on the hard tier, of which 55 items were used as training templates; 46.4% on the held-out half, where it is overconfident (ECE 0.263). The maker advises routing multi-step questions to a reasoning model and keeping a person in the loop for legal, medical, financial or safety decisions. Weights and code are MPL-2.0; the NOTICE credits Qwen3.5-4B-Base (Apache-2.0).

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability
- **classify** — assigns a category from a fixed taxonomy
- **route** — sends the input to one of several destinations

## At a glance

| | |
|---|---|
| Parameters | 2.7B |
| Base model | `Qwen/Qwen3.5-4B-Base` |
| Maker | Blackdrome AI Labs |
| Released | 2026-10-01 |
| License | mpl-2.0 |
| Reported accuracy | 82.6% |
| Reported latency | about 16 ms median per decision end to end on one H100 (15.1 ms p50 / 83.8 ms p95 on JevBench's public items) |

## Get the weights

```bash
pip install systemonemodels
systemone pull blackdrome-ai-labs/noma
```

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

## Read more

- [Model card](https://huggingface.co/BlackdromeAILabs/noma)
- [Code](https://github.com/blackdromeai-labs/noma)
- [Project page](https://blackdrome.tech/noma)

---

*This page was opened by System One for Blackdrome AI 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