Loom Spark 1.5 Flash

Experimental micro variant · Textile Labs

Naming note: "Flash" here means the same thing it does for Gemini Flash — a smaller, faster variant of the family, not a shrunk copy of the numbered model. This is a separate model trained from scratch, not a distillation of Loom Spark 1.5 (12.32M params). Loom Spark 1.5 Flash is 1.35M parameters — about a tenth the size.

Built in a deliberate under-2-hour experiment: how small can a Loom model get while keeping the two things that actually matter — staying in character, and never talking to itself. It knows almost nothing. That was the point.

The headline result

Every prior Loom release (v1, 1.5, 1.8) only learned <|endoftext|> at the very end of a whole training document. Mid-conversation, nothing told the model a turn had ended — so on any runtime without the exact right stop-token configuration, it would keep going and invent your next message itself. This happened to the founder testing 1.8 in Ollama the day it shipped.

Loom Spark 1.5 Flash's curriculum was rebuilt so <|endoftext|> follows every single reply, not just the end of a document. Verified: 442,333 / 442,333 model turns in the training corpus end in EOS. Tested with zero configuration — Ollama's stock chat template, no Modelfile, no stop tokens set by hand — across chat and raw completion endpoints: 0 self-dialogue turns. It stops because it learned to, not because a runtime told it to.

Honest limitations — read this first

At 1.35M parameters there is essentially no room left for facts once identity and conversational structure are learned. Expect:

• Wrong or garbled answers to almost any factual question. "Capital of France" may come back as "Buenos Aires." This isn't a bug — there simply isn't capacity left for a fact-core at this size.
• Offline <lookup> leakage on raw-model factual questions (18/30 measured, in line with the rest of the Loom line — see evaluation/RESULTS.md).
• Occasional cross-wiring between similar question types (e.g. answering "who are you" with the "who made you" response).
• What holds up well: identity, restraint, and emotional register — 0/12 identity probes leaked in testing, matching the best of the mainline models.

Prompt format

<tools:off>
<tools:off><user> who are you
<loom>

No trailing space after <loom> — see the mainline card for why that matters. With tools on, a reply may end in <lookup>query</lookup><|endoftext|>; your harness splices in <result>…</result> before continuing. Because of the EOS-every-turn fix, the model reliably stops even without a harness — the harness's <result> injection is still required for it to actually know anything found online, but nothing forces you to run one just to get a well-formed single reply.

Usage — transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("textilelabs/Loom-Spark-1.5-Flash")
model = AutoModelForCausalLM.from_pretrained("textilelabs/Loom-Spark-1.5-Flash")

prompt = "<tools:off>\n<tools:off><user> who are you\n<loom>"
ids = tok(prompt, return_tensors="pt", add_special_tokens=False).input_ids
out = model.generate(ids, max_new_tokens=100, do_sample=True,
                     temperature=0.8, top_k=50, pad_token_id=tok.eos_token_id)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=False))

Usage — Ollama

ollama create loom-spark-1.5-flash -f ollama/Modelfile
ollama run loom-spark-1.5-flash

Unusually for a base-style Loom model, this one is also safe to run with no Modelfile at all — ollama run hf.co/textilelabs/Loom-Spark-1.5-Flash will not talk to itself, though output quality is better with the correct template and stop tokens, which the Modelfile provides.

Usage — the agent harness

pip install ./harness
loom-chat --model textilelabs/Loom-Spark-1.5-Flash

Harness v0.2.2+ required — it auto-detects this model's per-turn marker format via an explicit flag in config.json rather than guessing from model size, since this model is far smaller than earlier heuristics assumed any Loom model would be.

Training

• Hardware: CPU-only Dell OptiPlex 9020, i5-4690, 4 cores, no GPU
• 2,625 steps, batch 32 × 256 tokens, ~34 minutes wall clock
• Corpus: same 70MB procedurally generated curriculum as Loom Spark 1.8, with EOS added after every model turn (not just document end)
• Final validation loss: 0.3544
• Architecture: 128d × 4 layers × 4 heads, vocab 4096 (fresh BPE, not shared with any other Loom generation)

Files

config.json / model.safetensors       transformers weights
tokenizer.json / tokenizer_config.json  4096-token custom BPE (Flash-specific)
loom-spark-1.5-flash-f32.gguf         GGUF for llama.cpp / Ollama
ollama/Modelfile                      correct template + stop tokens
harness/                              agent harness v0.2.2 with web search
evaluation/                           acceptance logs, self-termination proof

License

MIT. See LICENSE.