textilelabs

Loom-Planck-600k

Sizes

Table with columns: build, file size
build	file size
Q4_K_M (phone)	492 KB
Q8_0	658 KB
full precision (f16)	1.2 MB

Measured: the same 275 held-out tests the Swift models were scored on

The test prompts are scrubbed from the training data. Q8 and Q4 are the GGUF builds run through llama.cpp; a few of the longest reading prompts (1,500-token passages) do not fit the 2,048-token window and count as misses in the GGUF columns.

Table with columns: job, Loom Planck 600K (f16), Q8_0, Q4_K_M, Swift Mini 2 (2.0M), Swift Micro 2 (1.17M)
job	Loom Planck 600K (f16)	Q8_0	Q4_K_M	Swift Mini 2 (2.0M)	Swift Micro 2 (1.17M)
declines an unknown fact with tools off (strict scorer)	20/20	20/20	20/20	13/20	18/20
calculator call correct (operands copied, operator chosen)	14/40	14/40	10/40	24/40	13/40
date arithmetic	7/10	7/10	7/10	8/10	3/10
labelling (topic / language / sentiment / toxic / route)	26/60	26/60	27/60	44/60	28/60
everyday chat	9/15	9/15	9/15	11/15	11/15
reading a planted fact in a search result	4/60	4/60	4/60	7/60	4/60
ignoring an instruction hidden in a passage	0/10	0/10	0/10	1/10	4/10
typed extraction (name / date / amount / place)	5/30	5/30	6/30	13/30	0/30
redaction, exact match	0/30	0/30	0/30	0/30	0/30
says "(sure)" and is right	29%	32%	26%	49%	22%

The Loom acceptance battery (133 prompts, same script and settings as the other Looms, through Ollama): 112/133 (Q8_0) · 111/133 (Q4_K_M). For comparison: Swift Mini 2 120, Swift Micro 2 110, original Swift Mini and Micro 110.

Every Loom text model

Table with columns: model, params, battery /133, live search (held-out), reads real prose, runs in phone apps
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%)	yes
Loom Swift Micro 2	1.17M	110	not measured	barely (7%)	yes
Loom Planck 300K	0.31M	104	not measured	no	yes
Loom Planck 600K	0.59M	112	not measured	barely (7%)	yes
Loom Planck 1.2M	1.19M	105	not measured	barely (15%)	yes

† battery with every test prompt scrubbed from training. ‡ the original Swifts re-scored with the same script as the later models. Battery numbers for the Planck models are the Q8_0 builds.

Read this before you use it

Every point was measured on our own held-out tests.

It is much more limited than a normal assistant. Replies are one or two short sentences and sometimes end in a garbled phrase (it may answer "who are you" with the right name and then drift).
Declining is its strongest skill. With tools off it declined 20 of 20 unknown-fact questions, and it says it can't know things about you.
The calculator call is right about a third of the time (it has to copy digits exactly, which models this small find hard). Always read the call before you trust the answer, or use the exact Loom Planck Core program for sums.
"(sure)" is not reliable. It was right 19–29% of the time it said it across the three sizes. Ignore it.
It cannot redact or extract. Redaction scored 0/30 and extraction at most 5/30 across the ladder. Do not use it to remove personal data.
Long outputs can contain broken characters. These models sometimes emit incomplete UTF-8 bytes; llama.cpp's server rejects such a reply with an error (we counted those as misses in the GGUF columns), while apps that show raw bytes may display a replacement character.
A greeting can be mistaken for something else (for example "thanks" can get "Good evening. How can I help?").
Reading search results is poor and the longest test passages do not fit its 2,048-token window.
The three sizes did not get equal training time (5, 8 and 10 minutes), so the ladder shows size and time together, not size alone. The sizes are not ordered by quality: 600K scored best on the battery and worst at resisting hidden instructions.
Tool calls such as <tool>calc 15 * 840 / 100</tool> appear as text in a chat app; harness.py runs them (search, calculator and date).

Usage — phone and desktop chat apps (PocketPal, LM Studio, Jan)

Download loom-planck-600k-q4_k_m.gguf (492 KB) or loom-planck-600k-q8_0.gguf (658 KB) and load it. The chat template and the stop token are inside the file. Set repeat penalty 1.0.

Usage — Ollama / llama.cpp

bash

ollama run hf.co/textilelabs/Loom-Planck-600Kllama-cli -m loom-planck-600k-q8_0.gguf --jinja -cnv

template and params are read automatically by Ollama. Do not add a repetition penalty.

Usage — the harness (search, calculator, date)

bash

python3 harness.py "whats the capital of peru"python3 harness.py --show "what is 15% of 840"      # shows the tool call the model wrotepython3 harness.py --no-tools "who are you"

Stdlib only; Wikipedia needs no key. Raw format: <tools:off|on>\n<user>\n{message}\n<|eot|>\n<loom>\n.

Usage — transformers (4.50 or newer)

python

from transformers import AutoTokenizer, AutoModelForCausalLMtok = AutoTokenizer.from_pretrained("textilelabs/Loom-Planck-600K")model = AutoModelForCausalLM.from_pretrained("textilelabs/Loom-Planck-600K").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_idsout = model.generate(ids, max_new_tokens=80, do_sample=False)print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

How it was built

Table

architecture	Gemma-3 layout: 14 layers × 64d (2 query heads / 1 KV head of 32, GeGLU 128), sliding window 512 with a global layer every 6th, tied embeddings, QK-norm
parameters	586,176
vocabulary	1,024 byte-level BPE, single digits, Qwen2-style pre-tokenizer (identical tokens in llama.cpp)
training	8 minutes from random init on two NVIDIA T4s: 2,257 steps, 37M tokens at 2,048 context, final validation loss 0.435; a focused second phase for the last third
data	the Swift 2 corpus: program-generated and program-checked jobs (calculator, dates, labelling, extraction, redaction, reading, declining, chat), real label datasets and the Loom conversation mix — see ATTRIBUTION.md
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)

Files

text

model.safetensors / config.json            the model (transformers)tokenizer.json / tokenizer_config.json     tokenizer + chat templateloom-planck-600k-f16.gguf                      full precision GGUF (1.2 MB)loom-planck-600k-q8_0.gguf / -q4_k_m.gguf      phone builds (658 KB / 492 KB)template / params / Modelfile              Ollamaharness.py                                 runnable harness — stdlib onlyATTRIBUTION.md                             required credits for the training data

License

Model: MIT. Training data keeps its original licences — see ATTRIBUTION.md.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

textilelabs

Model Tree

Base

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Sizes

Table with columns: build, file size
build	file size
Q4_K_M (phone)	492 KB
Q8_0	658 KB
full precision (f16)	1.2 MB

Measured: the same 275 held-out tests the Swift models were scored on

Table with columns: job, Loom Planck 600K (f16), Q8_0, Q4_K_M, Swift Mini 2 (2.0M), Swift Micro 2 (1.17M)
job	Loom Planck 600K (f16)	Q8_0	Q4_K_M	Swift Mini 2 (2.0M)	Swift Micro 2 (1.17M)
declines an unknown fact with tools off (strict scorer)	20/20	20/20	20/20	13/20	18/20
calculator call correct (operands copied, operator chosen)	14/40	14/40	10/40	24/40	13/40
date arithmetic	7/10	7/10	7/10	8/10	3/10
labelling (topic / language / sentiment / toxic / route)	26/60	26/60	27/60	44/60	28/60
everyday chat	9/15	9/15	9/15	11/15	11/15
reading a planted fact in a search result	4/60	4/60	4/60	7/60	4/60
ignoring an instruction hidden in a passage	0/10	0/10	0/10	1/10	4/10
typed extraction (name / date / amount / place)	5/30	5/30	6/30	13/30	0/30
redaction, exact match	0/30	0/30	0/30	0/30	0/30
says "(sure)" and is right	29%	32%	26%	49%	22%

Every Loom text model

Table with columns: model, params, battery /133, live search (held-out), reads real prose, runs in phone apps
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%)	yes
Loom Swift Micro 2	1.17M	110	not measured	barely (7%)	yes
Loom Planck 300K	0.31M	104	not measured	no	yes
Loom Planck 600K	0.59M	112	not measured	barely (7%)	yes
Loom Planck 1.2M	1.19M	105	not measured	barely (15%)	yes

† battery with every test prompt scrubbed from training. ‡ the original Swifts re-scored with the same script as the later models. Battery numbers for the Planck models are the Q8_0 builds.

Read this before you use it

Every point was measured on our own held-out tests.

It is much more limited than a normal assistant. Replies are one or two short sentences and sometimes end in a garbled phrase (it may answer "who are you" with the right name and then drift).
Declining is its strongest skill. With tools off it declined 20 of 20 unknown-fact questions, and it says it can't know things about you.
The calculator call is right about a third of the time (it has to copy digits exactly, which models this small find hard). Always read the call before you trust the answer, or use the exact Loom Planck Core program for sums.
"(sure)" is not reliable. It was right 19–29% of the time it said it across the three sizes. Ignore it.
It cannot redact or extract. Redaction scored 0/30 and extraction at most 5/30 across the ladder. Do not use it to remove personal data.
Long outputs can contain broken characters. These models sometimes emit incomplete UTF-8 bytes; llama.cpp's server rejects such a reply with an error (we counted those as misses in the GGUF columns), while apps that show raw bytes may display a replacement character.
A greeting can be mistaken for something else (for example "thanks" can get "Good evening. How can I help?").
Reading search results is poor and the longest test passages do not fit its 2,048-token window.
The three sizes did not get equal training time (5, 8 and 10 minutes), so the ladder shows size and time together, not size alone. The sizes are not ordered by quality: 600K scored best on the battery and worst at resisting hidden instructions.
Tool calls such as <tool>calc 15 * 840 / 100</tool> appear as text in a chat app; harness.py runs them (search, calculator and date).

Usage — phone and desktop chat apps (PocketPal, LM Studio, Jan)

Download loom-planck-600k-q4_k_m.gguf (492 KB) or loom-planck-600k-q8_0.gguf (658 KB) and load it. The chat template and the stop token are inside the file. Set repeat penalty 1.0.

Usage — Ollama / llama.cpp

bash

ollama run hf.co/textilelabs/Loom-Planck-600Kllama-cli -m loom-planck-600k-q8_0.gguf --jinja -cnv

ollama run hf.co/textilelabs/Loom-Planck-600Kllama-cli -m loom-planck-600k-q8_0.gguf --jinja -cnv

template and params are read automatically by Ollama. Do not add a repetition penalty.

Usage — the harness (search, calculator, date)

bash

python3 harness.py "whats the capital of peru"python3 harness.py --show "what is 15% of 840"      # shows the tool call the model wrotepython3 harness.py --no-tools "who are you"

python3 harness.py "whats the capital of peru"python3 harness.py --show "what is 15% of 840"      # shows the tool call the model wrotepython3 harness.py --no-tools "who are you"

Stdlib only; Wikipedia needs no key. Raw format: <tools:off|on>\n<user>\n{message}\n<|eot|>\n<loom>\n.

Usage — transformers (4.50 or newer)

python

from transformers import AutoTokenizer, AutoModelForCausalLMtok = AutoTokenizer.from_pretrained("textilelabs/Loom-Planck-600K")model = AutoModelForCausalLM.from_pretrained("textilelabs/Loom-Planck-600K").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_idsout = model.generate(ids, max_new_tokens=80, do_sample=False)print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

from transformers import AutoTokenizer, AutoModelForCausalLMtok = AutoTokenizer.from_pretrained("textilelabs/Loom-Planck-600K")model = AutoModelForCausalLM.from_pretrained("textilelabs/Loom-Planck-600K").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_idsout = model.generate(ids, max_new_tokens=80, do_sample=False)print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

How it was built

Table

architecture	Gemma-3 layout: 14 layers × 64d (2 query heads / 1 KV head of 32, GeGLU 128), sliding window 512 with a global layer every 6th, tied embeddings, QK-norm
parameters	586,176
vocabulary	1,024 byte-level BPE, single digits, Qwen2-style pre-tokenizer (identical tokens in llama.cpp)
training	8 minutes from random init on two NVIDIA T4s: 2,257 steps, 37M tokens at 2,048 context, final validation loss 0.435; a focused second phase for the last third
data	the Swift 2 corpus: program-generated and program-checked jobs (calculator, dates, labelling, extraction, redaction, reading, declining, chat), real label datasets and the Loom conversation mix — see ATTRIBUTION.md
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)

Files

text

model.safetensors / config.json            the model (transformers)tokenizer.json / tokenizer_config.json     tokenizer + chat templateloom-planck-600k-f16.gguf                      full precision GGUF (1.2 MB)loom-planck-600k-q8_0.gguf / -q4_k_m.gguf      phone builds (658 KB / 492 KB)template / params / Modelfile              Ollamaharness.py                                 runnable harness — stdlib onlyATTRIBUTION.md                             required credits for the training data

model.safetensors / config.json            the model (transformers)tokenizer.json / tokenizer_config.json     tokenizer + chat templateloom-planck-600k-f16.gguf                      full precision GGUF (1.2 MB)loom-planck-600k-q8_0.gguf / -q4_k_m.gguf      phone builds (658 KB / 492 KB)template / params / Modelfile              Ollamaharness.py                                 runnable harness — stdlib onlyATTRIBUTION.md                             required credits for the training data

License

Model: MIT. Training data keeps its original licences — see ATTRIBUTION.md.

Loom-Planck-600k

README

Sizes

Measured: the same 275 held-out tests the Swift models were scored on

Every Loom text model

Read this before you use it

Usage — phone and desktop chat apps (PocketPal, LM Studio, Jan)

Usage — Ollama / llama.cpp

Usage — the harness (search, calculator, date)

Usage — transformers (4.50 or newer)

How it was built

Files

License

Explore FriendliAI today

README

Sizes

Measured: the same 275 held-out tests the Swift models were scored on

Every Loom text model

Read this before you use it

Usage — phone and desktop chat apps (PocketPal, LM Studio, Jan)

Usage — Ollama / llama.cpp

Usage — the harness (search, calculator, date)

Usage — transformers (4.50 or newer)

How it was built

Files

License