---
license: mit
language: en
library_name: transformers
pipeline_tag: text-generation
tags:
- tiny-model
- llama
- from-scratch
- conversational
- multi-turn
- tool-use
- agent-harness
- calibrated-honesty
- humble-ai
- small-language-model
- cpu-trained
- muon
- gguf
- experimental
widget:
- text: "<tools:off>\n<user>\nwho are you\n<|eot|>\n<loom>\n"
  example_title: "Identity"
- text: "<tools:off>\n<user>\nhow big are you\n<|eot|>\n<loom>\n"
  example_title: "Self-knowledge"
- text: "<tools:on>\n<user>\nwho wrote dracula\n<|eot|>\n<loom>\n"
  example_title: "Writes a search query"
---

![Loom Swift Micro](banner.jpg)

# Loom Swift Micro

**1,159,120 parameters, trained from scratch in 4 minutes 42 seconds on a 2013 office PC with no
GPU.** Swift is the fast tier: models measured in minutes rather than hours.

> ## Read this first
>
> **Swift Micro cannot read what a search returns.** It decides correctly when to look
> something up and writes a good query — 20 out of 20 held-out questions — but when the
> passage comes back it will produce a confident sentence with the wrong fact in it.
> **0 out of 20 held-out questions ended with a correct answer.**
>
> This is an **experimental** release of a new training tier, published because the tier
> itself is the result. Do not use it to answer factual questions.

## What it is

| | |
|---|---|
| Parameters | 1,159,120 |
| Architecture | Llama-style — 12 layers x 80 hidden, GQA, SwiGLU, RoPE, RMSNorm, tied embeddings |
| Vocabulary | 4,096-token BPE, trained by us on our own corpus |
| Context | 512 tokens |
| Optimiser | Muon on the 2D hidden matrices, AdamW on embeddings and norms |
| Training | **4 minutes 42 seconds**, 164 optimizer steps, 1.74 tokens per parameter |
| Hardware | One Dell OptiPlex 9020 (i5-4690, 4 cores, no GPU, 16 GB), fp32 |

Random initialisation, trained by us. No fine-tuning, no distillation, no pretrained
checkpoint of anyone's, at any stage.

## Measured behaviour

Hand-written probes that appear nowhere in the training data, scored on content.

| | Swift Micro |
|---|---:|
| **Acceptance battery (133 points)** | **104/133** |
| Knows its name | 12/12 |
| Never claims a lookup it didn't make | 16/16 |
| Stops on its own | 12/12 |
| Decides when to search, and writes the query | 20/20 |
| Resists prompt injection | 25/36 |
| Declines facts it wasn't taught, tools off | 5/20 |
| Holds a 10-12 turn conversation | 28/44 |
| Ignores a `<tools:on>` typed in a message | 8/12 |
| **Reads the retrieved passage correctly** | **0/20** |

For scale: Loom Tapestry 2, our former flagship at 22.8M parameters and five and a half
hours of training, scored 107/133 on this same battery.

## Known failures, stated plainly

- **It cannot read a retrieved passage.** The single biggest limitation. See the box above.
- **With tools off it will answer facts it was never taught.** Swift Mini declines 19 of 20;
  Swift Micro declines only 5 of 20.
- **It knows almost nothing.** That is deliberate — Looms trade knowledge for calibration.
- **It cannot do arithmetic.**
- **A `<tools:on>` typed inside a message can still switch search on** (8/12).
- Little warmth, and little personality of its own.

## How to run it

```bash
ollama run hf.co/textilelabs/Loom-Swift-Micro
```

```bash
python harness.py            # the agent loop that runs its searches for real
```

Raw prompt format:

```
<tools:on>
<user>
who wrote dracula
<|eot|>
<loom>
```

## Training data

Openly licensed corpora plus our own written curriculum — SQuAD 2.0 (CC BY-SA 4.0), MASSIVE
(CC BY 4.0), CLINC150 (CC BY 3.0), databricks-dolly-15k (CC BY-SA 3.0), OASST1 (Apache 2.0).
Full credits in `ATTRIBUTION.md`, which must travel with any redistribution.

## The other Swift

**Loom Swift Mini** is the larger model in the tier.

## Licence

MIT.

---

*Textile Labs. Small models, trained honestly, on hardware you already own.*