Loom 4.5

Loom Crucible 4.5

130M parameters Β· 32 layers Β· 2K context Β· runs on laptops and phones Β· Textile Labs

The Textile Labs woven logo threading itself together, then the words Loom 4.5

The first model of the Loom 4.5 family and the largest Loom trained in one go. Trained from scratch on 2 Γ— T4: randomly initialised weights, nothing fine-tuned from anyone else's checkpoint.

It is the most useful Loom so far. It does sums with a calculator (40 of 40), remembers what you told it earlier in the chat, writes the short email when you ask, searches for the actual subject, and passed every one of the 12 everyday checks that earlier Looms failed. It also says plainly when it can't know something, and never claims a lookup it didn't make.

What's new in 4.5

Loom Weave 4 Loom Crucible 4.5
size 55.6M 130M (32 layers Γ— 512)
calculator β€” <tool>calc …</tool> β†’ harness β†’ answer: 40/40
remembers what you said earlier in the chat 0/2 2/2
writes a short email or message when asked 0/2 ("I can't reach your phone or accounts") 2/2
"what is 2+2" (tools off Γ—2, tools on Γ—1) 0/3 ("That's 44") 3/3
quotes the sentence it used before answering β€” yes β€” in <think>
tool menus β€” the model quotes the line that fits; the harness maps it to the tool (25/45 right)
one run, seven sizes β€” trained with exits at layers 8/12/16/24, so smaller 4.5 models are cut from it (coming soon)

Measured

Every test prompt is scrubbed from the training data. Repeat penalty 1.0. Rows marked † are run through the transformers weights; the battery is also run on the Q4 phone file.

Everyday checks β€” 12 short conversations taken from real use (name memory, sums, emails, searching for the subject, rough typing, "what's my dog called", "what can you do"). Weave 4's name answer is counted as right for its own name:

Loom Weave 4 Loom Crucible 4.5
everyday checks 5/12 12/12

The acceptance battery, row by row:

row Loom Weave 4 Loom Crucible 4.5
A Β· says its own name 12/12 12/12
B Β· its own name under rough typing 12/12 12/12
C Β· 5-turn conversation stays on thread 5/5 5/5
D Β· answers from a search result 5/5 5/5
E Β· follow-up answered from the same result 4/5 2/5
F Β· says it looked, after a lookup 5/5 4/5
G Β· never claims a lookup it didn't make 16/16 16/16
H Β· admits what it can't know about you 8/8 8/8
I Β· says when a result doesn't contain the answer 1/5 1/5
J Β· never leaks a search tag with tools off 28/28 28/28
K Β· stops on its own 12/12 12/12
L Β· searches when it should, not for your private things 19/20 18/20
total 127/133 123/133 (Q4 phone file: 123/133 too)

Live search, end to end β€” 20 held-out everyday questions, live Wikipedia, the model writing its own query: 9/20 answered right (Weave 4: 10/20). It searched every time and wrote its own query (only 2 of 20 pasted the question); the right text came back 12 times and it read it correctly 9 times.

Held-out skill tests † (frozen before training):

Loom Crucible 4.5
calculator through the harness (40 mixed sums, percentages, division) 40/40
reading a short text and answering (2 / 5 / 12 / 45 sentences) 68% / 35% / 45% / 40%
reading natural text (60 questions, incl. 12 where the answer isn't there) 11/60 (Weave 4: 5/60)
uses what you told it over what it "knows" 19/30 (Weave 4: 2/30)
picks the right tool from a menu, via the harness's line matching 25/45
replies that stop by themselves 75%

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 β€” 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 no
Loom Weave 4 55.6M 127† 10/20 yes yes
Loom Crucible 4.5 130M 123† 9/20 yes β€” quotes its source yes

† scored on a battery with every test prompt scrubbed from training. Earlier rows are each model's release score.

Read this before you use it

Every point here was measured.

  • It is not the top scorer on our old quiz. 123/133 is below Weave 4 (127) and Crucible Preview (125). It lost points on follow-up questions about a search result (2/5) and on saying a result doesn't hold the answer (1/5). We released it because it is more useful in real conversations, not because it wins that test.
  • Reading is still the weak spot. On natural text it answers 11 of 60 correctly. It quotes a real sentence before answering, but often the wrong one, and it sometimes answers when the text doesn't say. Check the quote it gives.
  • "(sure)" is not a guarantee. Only 23 of its 75 "(sure)" reading answers were right.
  • Tool menus: don't trust the tool name it writes. It usually finds the right line but copies the name wrongly. Use pick_tool() in the harness, which matches the line it quoted (right tool 25/45) instead of the name it writes (17/45). Still check the choice before running anything that matters.
  • Long writing drifts. Emails and messages are short and on point; asked for something longer, it can repeat itself. A quarter of longer replies don't stop by themselves β€” keep num_predict at 256.
  • Rules from examples are weak (4 of 40 new combinations). It can follow a rule you state; it rarely infers one.
  • Live search: about half. 9 of 20 everyday questions answered right end to end; the harness's --show prints what it read.
  • No world knowledge to speak of. With tools off it will decline most fact questions β€” by design. Turn tools on.
  • Calculator, lookups and menus need the harness: the model writes the tag and stops.

Usage β€” phone and desktop chat apps (PocketPal, LM Studio, Jan)

Download loom-crucible-4.5-q4_k_m.gguf (81 MB) or loom-crucible-4.5-q8_0.gguf (139 MB) and load it. The chat template and the stop token are inside the file. Set repeat penalty 1.0.

Usage β€” the harness (search, calculator, tool menus, chat memory)

python3 harness.py "whats the capital of peru"
python3 harness.py "what's 15% of 1150"           # the calculator
python3 harness.py --show "who wrote hamlet"     # see what it searched and read
python3 harness.py                               # chat: it remembers the conversation; /new to start over

Tool menus from your own code:

from harness import pick_tool
name, raw = pick_tool("Put the heating up to 22", {
    "set_heating(temperature)": "set the heating to a temperature in degrees C",
    "lock_door()": "lock the front door"})

Stdlib only; Wikipedia needs no key. Never feed a failed lookup back as a result β€” the model answers from the error text.

Usage β€” Ollama / llama.cpp

ollama run hf.co/textilelabs/Loom-Crucible-4.5 "who are you"
llama-cli -m loom-crucible-4.5-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 β€” transformers

from transformers import AutoTokenizer, AutoModelForCausalLM
tok = AutoTokenizer.from_pretrained("textilelabs/Loom-Crucible-4.5")
model = AutoModelForCausalLM.from_pretrained("textilelabs/Loom-Crucible-4.5").eval()
prompt = "<tools:off>\n<user>\nwho are you\n<|eot|>\n<loom>\n"
ids = tok(prompt, return_tensors="pt", add_special_tokens=False).input_ids
out = model.generate(ids, max_new_tokens=160, do_sample=False, eos_token_id=0, pad_token_id=1)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

Raw format: <tools:off>\n<user>\n{message}\n<|eot|>\n<loom>\n (or <tools:on> to allow search and the calculator).

How it was built

architecture Llama: 32 layers Γ— 512, FFN 2,048, GQA 8 heads / 2 KV, RoPE, tied embeddings
parameters 130,056,704
vocabulary 16,384 BPE, single digits
run 1 8.2 h from random init, 445M tokens: reading, skills, replies, Loom behaviour and general text at 1K, then 2K. An extra loss at layers 8/12/16/24 so smaller models can be cut out of it
run 2 1.5 h on its own weights, 86M tokens: follow-ups, tool menus, memory, short emails, sums, searching for the subject β€” each taught with its contrast, replayed with the run-1 data
released weights a 30/70 average of run 1's best save and run 2 β€” keeps run 2's new skills and most of run 1's reading
optimiser Muon on hidden matrices, AdamW on embeddings and norms; per-conversation attention masking
hardware Kaggle, 2 Γ— NVIDIA T4

Files

model.safetensors / config.json            the model (transformers)
tokenizer.json / tokenizer_config.json     tokenizer
loom-crucible-4.5-f16.gguf                 full precision GGUF (261 MB)
loom-crucible-4.5-q8_0.gguf / -q4_k_m.gguf phone builds (139 / 81 MB)
template / params / Modelfile              Ollama
harness.py                                 search + calculator + tool menus + chat memory β€” stdlib only
banner.jpg / logo.jpg / launch.gif         images
ATTRIBUTION.md                             required credits for the training data

License

Model: MIT. Training data keeps its original licences β€” see ATTRIBUTION.md.

Downloads last month
58
Safetensors
Model size
0.1B params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including textilelabs/Loom-Crucible-4.5