A compact, bilingual, vision-capable model specialized for Canadian legal and regulatory work — in English and Canadian French.
-orange){kind=icon}
flash-1-mini is a 4.54B-parameter fine-tune of Qwen3.5-4B trained by Alpine Pacific Trading Inc. (operating as SimpleDirect®) for Canadian legal and regulatory workflows. It is strong at correctly-formatted legal citations and instruction-following, bilingual in English and Canadian French, and covers both common law and Québec civil law traditions. It retains the base model's general reasoning and vision capabilities.
This repository contains code and documentation only. Model weights live on Hugging Face: huggingface.co/simpledirect/flash-1-mini
Numbers are flash-1-mini vs. Qwen3.5-4B (base), measured under identical conditions. Full methodology and the CBLRE eval-set documentation are in the Hugging Face model card.
| Metric | flash-1-mini | Base (Qwen3.5-4B) | Delta |
|---|---|---|---|
| Legal citation integrity (CBLRE citation) | 42.1% | 15.8% | +26.3 pp |
| Instruction-following (IFEval prompt-strict) | 53.2% | 30.3% | +22.9 pp |
| Bilingual parity (privacy-compliance EN/FR) | 90.9% / 90.9% | — | ratio 1.00 |
| English legal — international law (MMLU) | 76.0% | 70.3% | +5.7 pp |
| General knowledge (MMLU) | ~69.8% | ~69.8% | ≈ 0 |
| Complex reasoning (BBH) | 79.0% | 68.6% | +10.4 pp |
| Vision-capable (image + text) | Yes | Yes | — |
Key takeaways:
- 2.7× more reliable legal citations (42.1% vs 15.8%)
- +22.9 points instruction-following — more likely to comply with format and scope constraints
- Bilingual parity ratio 1.00 — identical privacy-compliance accuracy in English and French
- General knowledge unchanged; complex reasoning substantially improved
Note on licenses: This repository (code, examples, documentation) is MIT. The flash-1-mini model weights, hosted on Hugging Face, are Apache 2.0 (as a derivative of Qwen3.5-4B). When citing the model, use Apache 2.0.
Honest disclosure of verified regressions vs. the base model (identical conditions):
| Metric | flash-1-mini | Base (Qwen3.5-4B) | Delta |
|---|---|---|---|
| Retrieval / RAG | 75.5% | 80.5% | −5.0 pp |
| Function-calling (BFCL v4) | 28.6% | 37.7% | −9.1 pp |
| French professional-law MCQ (Global-MMLU FR) | 44.6% | 49.0% | −4.4 pp |
| CBLRE Québec civil law | 90.0% | 95.0% | −5.0 pp |
If your workflow is primarily retrieval-augmented generation, tool-calling, French professional-law MCQ, or Québec civil law, benchmark carefully before deploying over the base model.
pip install "transformers>=5.5" torch accelerate pillowimport torch
from transformers import AutoModelForImageTextToText, AutoProcessor
processor = AutoProcessor.from_pretrained("simpledirect/flash-1-mini")
model = AutoModelForImageTextToText.from_pretrained(
"simpledirect/flash-1-mini",
torch_dtype=torch.bfloat16,
device_map="auto",
# no trust_remote_code needed
)
messages = [
{
"role": "user",
"content": [{"type": "text", "text": "Summarise the key protections under the Canada Labour Code Part III."}],
}
]
# enable_thinking=False passed directly to apply_chat_template (not via chat_template_kwargs)
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = processor(text=text, return_tensors="pt").to(model.device)
with torch.no_grad():
output_ids = model.generate(**inputs, max_new_tokens=512, do_sample=False, temperature=0)
new_ids = output_ids[0][inputs["input_ids"].shape[-1]:]
print(processor.decode(new_ids, skip_special_tokens=True))See examples/quickstart_transformers.py for English + French
prompts and a thinking-on example.
# Download
huggingface-cli download simpledirect/flash-1-mini \
gguf/flash-1-mini-20260602-Q5_K_M.gguf --local-dir ./flash-1-mini-gguf
# Ollama
ollama create flash-1-mini -f Modelfile
ollama run flash-1-mini "What are the notice requirements under the Canada Labour Code?"See examples/gguf_llamacpp.md and
examples/ollama.md for full instructions.
| Variant | Modalities | Size | Location |
|---|---|---|---|
| bf16 (full weights) | Text + image | ~9 GB | HF main repo |
| Q6_K GGUF | Text only | ~3.3 GB | gguf/ folder |
| Q5_K_M GGUF | Text only | ~2.9 GB | gguf/ folder |
| Q4_K_M GGUF | Text only | ~2.6 GB | gguf/ folder |
GGUFs are text-only — they do not include the vision encoder. For image + text inference, use the bf16 weights.
The model thinks by default, emitting <think>…</think> before its visible answer.
For production legal use, disable thinking for deterministic output:
# Direct kwarg — do NOT wrap in chat_template_kwargs
text = processor.apply_chat_template(messages, ..., enable_thinking=False)Thinking ON is useful for complex multi-step reasoning (e.g., cross-jurisdictional analysis).
| Property | Value |
|---|---|
| Architecture | Qwen3_5ForConditionalGeneration |
| Parameters | 4.54B |
| dtype | bfloat16 |
| Hidden size | 2560 |
| Layers | 32 |
| Attention heads | 16 |
| Vocabulary | 248 320 |
| Context length | 262 144 |
| Tied embeddings | Yes |
| Transformers requirement | >= 5.5 (native; no trust_remote_code) |
flash-1-mini is designed as a drafting and research assistant for Canadian legal and regulatory workflows — summarisation, citation lookup, bilingual Q&A, document analysis, and compliance checking.
It assists; it does not replace professional judgment.
- Always verify citations against primary sources before relying on them. Citation accuracy improved substantially over base, but is not perfect.
- This model does not provide legal advice and should not be presented to clients as doing so.
- Outputs may contain errors, outdated information, or jurisdiction mismatches. Human review is required before any use in professional or legal contexts.
Repository code (this repo): MIT — Copyright © 2026 Alpine Pacific Trading Inc. (operating as SimpleDirect®)
Model weights: Apache-2.0 — derivative of Qwen/Qwen3.5-4B. See the model's LICENSE and NOTICE on Hugging Face.
Full benchmark methodology, evaluation harness details, and CBLRE eval-set documentation are available in the Hugging Face model card. Training code, data, and hyperparameter recipes are intentionally not published.