# Tamkimd: tamev

> Small encoder decision models: a state, a typed question and your option list in, a probability per option out, in one forward pass on CPU. choice, noul and score share one call. Headline Micro tier is 22.8M params on MiniLM; the maker says accuracy holds only in-distribution.

- Page: https://systemonemodels.ai/tamkimd/tamev
- API: https://api.systemonemodels.ai/v1/models/tamkimd/tamev
- Download: `pip install systemonemodels && systemone pull tamkimd/tamev`

## Facts

| | |
|---|---|
| Maker | Tamkimd (https://systemonemodels.ai/tamkimd) |
| Decides | choice, score, noul |
| Architecture | tamev |
| Base model | sentence-transformers/all-minilm-l6-v2 |
| Parameters | 23M |
| Context | 128 tokens |
| Licence | apache-2.0 |
| Availability | Open weights |
| Released | 2026-09-25 |
| Latest version | 2026.09 |

## Reported evaluation

Suite: TAMEV held-out set (19,999 typed-decision items re-formed from public datasets; in-distribution, the maker's own). Numbers are the publisher's own.

- Decision accuracy: 82.2%
- Calibration error (ECE): 0.022
- Median latency: 7.3 ms
- p95 latency: 13.1 ms

## Model card

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

# TAMEV-Micro-MiniLM

Small encoder decision models: a state, a typed question and your option list in, a probability per option out, in one forward pass on CPU. choice, noul and score share one call. Headline Micro tier is 22.8M params on MiniLM; the maker says accuracy holds only in-distribution.

A decoupled bi-encoder with a bilinear pointer head shared across option slots, so option order cannot change the answer; options are passed per call. The maker warns that top-1 is near chance outside the training distribution, so fine-tune before use; the published numbers are on an in-distribution held-out set. Weights have been replaced several times, so pin a revision; trust_remote_code is required. The training data includes Yelp and Amazon reviews, which carry upstream terms. Siblings nano-tinybert, nano-flash-tinybert and small-modernbert; larger Qwen tiers are announced without weights. A TorchScript file in the repo is flagged by the Hub scanner and is not listed here.

## 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

## At a glance

| | |
|---|---|
| Parameters | 22.8M |
| Base model | `sentence-transformers/all-MiniLM-L6-v2` |
| Maker | Tamkimd |
| Released | 2026-09-25 |
| License | apache-2.0 |
| Reported accuracy | 82.2% |
| Reported latency | 7.29 ms p50 / 13.07 ms p95 on one CPU thread on an Apple M4, warm option cache |

## Get the weights

```bash
pip install systemonemodels
systemone pull tamkimd/tamev
```

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

## Read more

- [Model card](https://huggingface.co/Tamkimd/tamev-micro-minilm)
- [Source code](https://github.com/tamkimd/tamev)

---

*This page was opened by System One for Tamkimd, 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