--- license: mit language: en library_name: transformers pipeline_tag: text-generation tags: - tiny-model - gemma3 - from-scratch - conversational - tool-use - agent-harness - calculator - humble-ai - small-language-model - muon - gguf - on-device - experimental widget: - text: "\n\nwho are you\n<|eot|>\n\n" example_title: "Identity" - text: "\n\nwhat's 482 + 539?\n<|eot|>\n\n" example_title: "Writes a calculator call" - text: "\n\nwhat's the weather in Paris right now?\n<|eot|>\n\n" example_title: "Limits" ---
Loom Swift Micro 2
# Loom Swift Micro 2 **1,172,832 parameters · 12 layers × 96d · 2K context · runs on phones · Textile Labs** Loom Swift Micro 2 is a 1.17-million-parameter model trained for five minutes. It is the smallest Loom: it chats briefly, knows who it is, declines what it can't know, and writes calculator and date calls — and it runs in phone apps. It is the second generation of [Loom Swift Micro](https://huggingface.co/textilelabs/Loom-Swift-Micro), retrained on Kaggle's two T4 GPUs for the same 5 minutes the original spent on a CPU, and rebuilt on the Gemma-3 layout so it runs in PocketPal, LM Studio, Jan, llama.cpp and Ollama straight from the GGUF. Trained from scratch: randomly initialised weights, nothing fine-tuned from anyone's checkpoint. > ## Read this first > This is a **tiny model for tiny jobs**, and it is honest about the limits we measured. It is not a reader and it is not a reasoner. > What it does is: say who it is, decide when a tool is needed, write a calculator / date / search call, and (Mini) label and extract from short text. > It is published as a small, fast, on-device experiment — not as a strict upgrade. The "Read this before you use it" section below lists every miss. ## What changed from the original | | Loom Swift Micro | **Loom Swift Micro 2** | |---|---|---| | layout | Llama | **Gemma-3 style** (local sliding window + a global layer every 6th) | | tokenizer | 4,096 BPE | **4,096 BPE rebuilt, every digit its own token, identical in llama.cpp** | | context | 512 | **2,048** | | runs in phone/desktop chat apps | no | **yes — chat template and end-of-turn built into the GGUF** | | new jobs | — | calculator calls, date arithmetic, labelling, extraction, redaction, reading a search result | | trained on | a 2013 office CPU | Kaggle 2 × T4 | | training time | 10 / 5 minutes | the same: **5 minutes** of training | ## Measured against the original Same tests, same scripts, same settings (repeat penalty 1.0), both through Ollama. Every test prompt is scrubbed from the training data. **The Loom acceptance battery**, row by row: | row | original | **Loom Swift Micro 2** | |---|---:|---:| | A · says its own name | 11/12 | 11/12 | | B · its own name under rough typing | 11/12 | 11/12 | | C · 5-turn conversation stays on thread | 5/5 | 5/5 | | D · answers from a search result | 0/5 | 3/5 | | E · follow-up answered from the same result | 0/5 | 0/5 | | F · says it looked, after a lookup | 5/5 | 0/5 | | G · **never** claims a lookup it didn't make | 16/16 | 16/16 | | H · admits what it can't know about you | 7/8 | 5/8 | | I · says when a result doesn't contain the answer | 0/5 | 3/5 | | J · never leaks a search tag with tools off | 28/28 | 28/28 | | K · stops on its own | 12/12 | 12/12 | | L · searches when it should | 15/20 | 16/20 | | **total** | **110/133** | **110/133** | **New jobs** (our own frozen test set; the original Swifts never trained on most of these, so most of their zeros are expected): | job (275 held-out items, each prompt scrubbed from training) | original Swift Micro | **Swift Micro 2** | |---|---:|---:| | everyday chat | 4/15 | **11/15** | | declining an unknown fact, tools off (strict scorer) | 18/20* | 18/20 | | ignoring an instruction inside a passage | 0/10 | **4/10** | | labelling (topic / language / sentiment / toxic / route) | 0/60 | **28/60** | | calculator call correct | 0/40 | **13/40** | | date arithmetic | 0/10 | **3/10** | | reading a retrieved passage | 0/60 | **4/60** | | typed extraction | 0/30 | 0/30 | | redaction (exact match) | 0/30 | 0/30 | | saying "(sure)" and being right | — | 22% of 82 | *the original was scored with a looser scorer. Q8_0 and Q4_K_M behave the same within a few items (Q8: label 30/60, inject 5/10; Q4: label 27/60, inject 1/10 — injection resistance is the least stable part). ## Every Loom text model | model | params | battery /133 | live search (held-out) | reads real prose | runs in phone apps | |---|---:|---:|---:|---|---| | Loom Spark 2 | 19.9M | ~97 | 2/20 | no | no | | Loom Tapestry 2 | 22.8M | 107 | — | curated only | no | | Loom Tapestry 3 Flash | 7.18M | 112 | 3/20 | curated only | no | | Loom Spark 3 Flash | 7.18M | 119 | 5/20 | curated only | no | | Loom Spark 3 | 12.2M | 120 | 7/20 | curated only | no | | Loom Spark 3.2 | 22.8M | 122† | 7/20 | curated only | no | | Loom Weave 3 | 31.5M | 120 | 6/20 | yes | no | | Loom Tapestry 3 | 69.2M | 123 | 12/20 | yes + multi-hop | no | | Loom Crucible Preview | 155.0M | 125† | 10/20 | yes — best curated reader | no | | Loom Weave 4 | 55.6M | 127† | 10/20 | yes | yes | | Loom Swift Mini (original) | 2.0M | 110‡ | — | no | no | | Loom Swift Micro (original) | 1.16M | 110‡ | — | no | no | | Loom Swift Mini 2 | 2.0M | 120 | not measured | barely (12% on 200–1,500-token passages) | yes | | **Loom Swift Micro 2** | **1.17M** | **110** | not measured | barely (7%) | **yes** | † scored on a battery with every test prompt scrubbed from training. ‡ the two original Swifts re-scored today with the same script and settings as the Swift 2s (their release scores were 106 and 104). Earlier rows are each model's release score. Never compare scores from different command lines. ## Read this before you use it Every point was measured on our own held-out tests. - **It is weaker than Swift Mini 2 at almost everything.** Calculator calls 13/40, dates 3/10, labelling 28/60, reading 4/60, extraction and redaction 0. - **On the acceptance battery it ties the original (110/133) and is worse in two places**: attribution after a lookup went from 5/5 to 0/5 and admitting an unknowable from 7/8 to 5/8. It is better at reading a result (3/5 vs 0/5) and at saying a result doesn't contain the answer (3/5 vs 0/5). - **Its "(sure)" is nearly meaningless**: right 22% of the time it says it. - **Do not use it for calculation, extraction or redaction.** It is here for what a 1-million-parameter file can do in five minutes of training: answer who it is, chat briefly, decline, and decide when a tool is needed. - Replies are one or two short sentences. 2,048 tokens of context. ## Usage — phone and desktop chat apps (PocketPal, LM Studio, Jan) Download `loom-swift-micro-2-q4_k_m.gguf` (1.2 MB) or `loom-swift-micro-2-q8_0.gguf` (1.4 MB) and load it. The chat template and the stop token are inside the file. Set **repeat penalty 1.0**. The model only writes tool calls (`calc 482 + 539`); a chat app will not run them for you — use the harness for that. ## Usage — the harness (search, calculator, date) ```bash python3 harness.py "whats the capital of peru" python3 harness.py "what is 482 + 539" python3 harness.py --show "who wrote hamlet" # see what it searched and read python3 harness.py --no-tools "who are you" ``` Stdlib only; Wikipedia needs no key. Calculator and date calls run locally; a calculator error comes back as `error`. **Never feed a failed lookup back as a result** — the model answers from the error text. ## Usage — Ollama / llama.cpp ```bash ollama run hf.co/textilelabs/Loom-Swift-Micro-2 "who are you" llama-cli -m loom-swift-micro-2-q8_0.gguf --jinja -cnv # uses the template in the GGUF ``` `template` and `params` are read automatically by Ollama. **Do not add a repetition penalty.** ## Usage — transformers (4.50 or newer) ```python from transformers import AutoTokenizer, AutoModelForCausalLM tok = AutoTokenizer.from_pretrained("textilelabs/Loom-Swift-Micro-2") model = AutoModelForCausalLM.from_pretrained("textilelabs/Loom-Swift-Micro-2").eval() prompt = tok.apply_chat_template([{"role": "user", "content": "who are you"}], tokenize=False, add_generation_prompt=True) ids = tok(prompt, return_tensors="pt", add_special_tokens=False).input_ids out = model.generate(ids, max_new_tokens=120, do_sample=False) print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True)) ``` Raw format (tools on or off): `\n\n{message}\n<|eot|>\n\n`. ## How it was built | | | |---|---| | architecture | Gemma-3 layout: 12 layers × 96d (10 sliding-window 512 + 2 global), 2 query heads / 1 KV head of 32, GeGLU 160, tied embeddings, QK-norm | | parameters | 1,172,832 | | vocabulary | 4,096 byte-level BPE, single digits, Qwen2-style pre-tokenizer (identical tokens in llama.cpp) | | training | 5 minutes from random init, 608 steps, 40M tokens at 2,048 context; a second, focused phase for the last third | | optimiser | Muon on 2-D hidden matrices, AdamW on embeddings and norms; per-conversation attention masking | | attention check | before the run, the trainer's attention is checked against the reference implementation (loss and gradients identical) | | hardware | Kaggle, 2 × NVIDIA T4 | ## Files ``` model.safetensors / config.json the model (transformers) tokenizer.json / tokenizer_config.json tokenizer + chat template loom-swift-micro-2-f16.gguf full precision GGUF (2.5 MB) loom-swift-micro-2-q8_0.gguf / -q4_k_m.gguf phone builds (1.4 MB / 1.2 MB) template / params / Modelfile Ollama harness.py runnable harness — search, calculator, date; stdlib only ATTRIBUTION.md required credits for the training data ``` ## License Model: MIT. Training data keeps its original licences — see ATTRIBUTION.md.