Available now · Apache 2.0
flash-1-mini
The bilingual legal AI that gets the citation right.
4 billion parameters. Open weights. Bilingual English and French. It runs on a laptop — and once you've downloaded it, it's yours. No API key, no account, nothing anyone can revoke.
What it does
A specialist, not a generalist
flash-1-mini is tuned for the two things that decide whether a legal or compliance team can trust a model's output: getting the citation right, and following the instruction exactly. It's still a capable general model — but this is what it's built for.
Legal citation accuracy
Produces correctly-formatted Canadian legal citations 2.7× more reliably than the base it's built on — 42.1% vs 15.8% on the CBLRE benchmark.
Instruction-following
Follows detailed, multi-part instructions 22.9 points more accurately than the base on IFEval — 53.2% vs 30.3%.
Bilingual & vision-capable
Works in English and French with privacy parity across both, and reads images — vision is inherited intact from the base model.
The numbers
Verified benchmarks
Every number is measured against the Qwen base flash-1-mini is fine-tuned from, under identical conditions, with the same scorer. All of it reproduces from the public benchmarking methodology and the CBLRE evaluation suite.
| Capability | Base | flash-1-mini | Δ |
|---|---|---|---|
| Legal citation integrity (CBLRE) | 15.8% | 42.1% | +26.3 (2.7×) |
| Instruction-following (IFEval) | 30.3% | 53.2% | +22.9 |
| English legal — international law (MMLU) | 70.3% | 76.0% | +5.8 |
| English legal — jurisprudence (MMLU) | 79.6% | 81.5% | +1.9 |
| Complex reasoning (BBH) | 68.6% | 79.0% | +10.4 |
| TruthfulQA MC2 | 48.8% | 50.8% | +2.0 |
| General knowledge (MMLU) | 69.8% | 69.8% | preserved |
| Bilingual privacy parity (FR/EN, CBLRE) | — | 1.00 | parity |
Honest tradeoffs
A specialist gives up ground it isn't built to hold. flash-1-mini trades some retrieval and tool-use performance for the gains above. We publish that here, at the same weight as the wins — not buried in a footnote.
| Capability | Base | flash-1-mini | Δ |
|---|---|---|---|
| Retrieval-grounded QA (Canadian RAG) | 80.5% | 75.5% | −5.0 |
| Function-calling (BFCL v4) | 37.7% | 28.6% | −9.1 |
| French professional law (Global-MMLU FR) | 49.0% | 44.6% | −4.4 |
| CBLRE Quebec civil law | 95.0% | 90.0% | −5.0 |
For retrieval-heavy (RAG) or tool-use-heavy agentic workflows, the base model or a different specialization may serve you better. We'd rather tell you that than oversell.
Results from checkpoint flash-1-mini-20260602.
Who it's for
For people who can't get the citation wrong
Legal & compliance teams
Drafting and research where a wrong citation is a real problem — in English or French.
Policy & regulatory analysts
Canadian regulated workflows spanning common law and Quebec civil law.
Edge & air-gapped deployments
Runs where cloud APIs can't — from Raspberry Pi-class hardware to a small server, fully offline.
Evaluators & researchers
Reproduce every number with the open methodology and the CBLRE suite. Audit it yourself.
How to install
Three ways to run it
Open weights — download the file and point your runtime at it. No account, no key.
Python · transformers
from transformers import AutoProcessor, AutoModelForCausalLM
import torch
model = AutoModelForCausalLM.from_pretrained(
"simpledirect/flash-1-mini",
torch_dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True,
)
processor = AutoProcessor.from_pretrained("simpledirect/flash-1-mini")Production serving · vLLM
vllm serve simpledirect/flash-1-mini \
--dtype bfloat16 \
--trust-remote-code \
--max-model-len 8192Laptop & edge · llama.cpp / Ollama / LM Studio
# GGUF builds: simpledirect/flash-1-mini-gguf
ollama run hf.co/simpledirect/flash-1-mini-gguf:Q4_K_MGGUF quantizations: Q4_K_M, Q5_K_M, Q8_0. Use Q4_K_M for laptop and Pi-class hardware; Q8_0 for a desktop or server with 16 GB+ of RAM.
Open alongside the model
Four public goods
We're releasing the evaluation infrastructure with the model — links go live as each piece lands. The standard is the point.
Canadian Bilingual Legal Corpus
The dataset flash-1-mini is fine-tuned on, with provenance documentation.
Coming soonCBLRE Evaluation Suite (Preview)
Canadian Bilingual Legal & Regulatory Evaluation — six tracks, bilingual ground truth, reproducible scoring.
Coming soonCanadian AI Evaluation Methodology v1.0
How to evaluate AI for Canadian regulated workflows.
Coming soonModel Benchmarking Methodology v1.0
How we measured what we measured — the reproducibility protocol.
Coming soonApache 2.0. You own the file.
flash-1-mini is released under the Apache License 2.0 — the merged weights, the adapter, the code, and the documentation. It is a derivative work of an open, Apache-2.0-licensed Qwen base model, and we attribute that base as the license requires (see the NOTICE file). Apache 2.0 means commercial use, redistribution, and modification are all fine. No employee-count limits. Nothing to revoke.
Cite it
@misc{simpledirect2026flash1mini,
title = {flash-1-mini: A Bilingual Canadian Legal AI},
author = {Naik, Ayush and Pu, George and Grover, Vikas},
organization = {SimpleDirect / Alpine Pacific Trading Inc.},
year = {2026},
url = {https://huggingface.co/simpledirect/flash-1-mini}
}Who built it
A five-person team in Toronto
George Pu (Founder), Vikas Grover (COO & CFO), and Ayush Naik (Engineering), plus a small bench of engineering contractors. Built in Toronto, Canada.
Have a question, or a procurement inquiry? Contact us
AI you can own.
The full story behind flash-1-mini — why a bilingual Canadian legal model, and what it means to actually own one — is on Founder Reality.
Read the launch essay