AI & AGI
Bilingual by Design: Building a Small Canadian LLM on One Card
A 27B open model shrunk to about 9 GB, built and tested in English and French, running in stock llama.cpp on one workstation card. v1 is out tod…

Building custom LLMs has been a learning journey. The success has been shrinking a 27 billion parameter model to about 9 GB, so it runs on one workstation card with no cloud involved, and the results are a couple of well documented LLM builds. At some point I went onto the socials and said, almost as an aside, "English and French next, because Canada".
That aside turned out to be the whole point.
Part of the reason was VoxRelay, a voice AI assistant that books appointments by phone. It runs on my own GPU compute, with speech-to-speech times in the range of larger competitors like ElevenLabs. Measured on my own box (warm, English and French): speech-to-text in about 63 ms, the language model's first token in about 40 ms, and the first sound of the reply in about 145 ms. That's roughly a quarter of a second of processing from the caller finishing to the reply starting. It was the Spring / Summer of 2026, and big-card GPU compute felt approachable.
In building and testing it, I kept running into the same worry: when it spoke French to a Canadian caller, was it getting it right? The available models were trained mostly on European French. Were the words, the expressions and the tone a little off for someone in Montréal or Québec City? Would people trust it, or would it come across as careless, even insensitive, to the people you most want to serve? VoxRelay's audience is seniors and voice-first people who prefer a phone call to a screen notification. Rural landlines, where the local small business still calls its customers because they're not online.
That's the UX consultant in me, trying to solve a problem of trust and authenticity. When I looked for a small model built and tested for Canadian French, I couldn't find one that met the need. That gap was the inspiration for this build, and for releasing it as a v1 anyone can use.
Why a bilingual build
Most small AI models are built and tested in English, because the biggest tech markets transact in English. In Canada, that leaves French with whatever is left over.
Shrinking a model makes this worse. The compression decides which parts of the model matter most by watching it work. If it only watches English, French pays the price.
So this build was calibrated on English and French and tested in both languages at every step. It is a ~9 GB version of the open Qwen3.8 27B model, and it runs in stock llama.cpp with no forks and no special engine. Every build and test so far ran on a single workstation card.
Results
The comparison is against the full-size original and the best public build at the same size, Unsloth's UD-Q2_K_XL.
How closely the small model tracks the original:
| English | French | Québec French | |
|---|---|---|---|
| Standard llama.cpp build at this size | well behind | about 2× more drift than this build | about 2× more drift than this build |
| Unsloth UD-Q2_K_XL | best | level | level |
| This build | close, still behind | level, a hair ahead | level, a hair ahead |
Task benchmarks, run through the open engine the way a user would run it:
| Benchmark | Original (54 GB) |
This build (9 GB) |
Unsloth (9 GB) |
|---|---|---|---|
| Reading comprehension, French (Belebele) | 94.9% | 94.7% | 93.6% |
| Reading comprehension, English (Belebele) | 96.2% | 96.1% | 95.9% |
| Maths asked in French (MGSM-fr) | 88.4% | 89.2% | 88.8% |
| Maths asked in English (MGSM) | 95.6% | 96.4% | 95.6% |
| Maths, English (GSM8K) | 96.4% | 95.4% † | 95.3% |
| Following instructions, English (IFEval) | 82.1% | 80.6% | 83.0% |
| Following instructions, French (IFEval-fr) | 73.8% | 72.6% | 73.2% |
| Québec French grammar (COLE QFrCoLA) | 80.0% | 76.2% | 75.6% |
† Earlier builds sometimes returned an empty answer on this test, up to one in eleven. The cause was how the model learned to end its turn. In v1, 2 of 1,319 answers come back empty.
What testing caught
- Empty answers on the maths test, above. Found, traced and fixed before release.
- A civic error. The model called Québec "the bilingual province" in French. It's New Brunswick; French is Québec's only official language. The full-size original also got this wrong when asked a few different ways. A small set of civic facts in both languages, each checked against the law, went into training. v1 answers correctly in both languages.
- The amount of Québec text matters. Most French on the web is European French, so a modest amount of public Québec text went into training, with a separate Québec judge to measure it. A small dose helped across the board, including on the empty answers. Twice as much made results worse; the reason is still open.
The stretch goal
- Accurate in both languages. Close the English gap and keep growing Canadian French.
- Consistent. The same question gets the same answer every time.
- Measured and published. Every claim backed by a number, with an evaluation set others can rerun.
- Small enough to own. One card, stock open tooling, nothing leaving the building.
Why this matters for Canada
Federal institutions have to serve people equally in English and French, and the Treasury Board's guidance on generative AI asks that output quality be checked in each official language. In Québec, the province's guidance for public bodies leans toward open models, data kept in Québec and systems run on their own servers, and Bill 96 expects French versions of software to work as well as the English ones.
Others did the groundwork. Université Laval's COLE benchmark measures how well models understand Québec French. The QuebecLLM project from Concordia and Laval released the first open Québec French base models. COLE's tests also show open instruct models like Chocolatine doing well in French. This build adds Québec French scores taken on the exact file you download, with every answer published so anyone can check them.
Use it
v1 is a starting point, not a finished product. It's free to download and run.
- Download: Hugging Face
- Run it:
ollama run vveerrgg/qwen3.8-27b-bilingual-canadian - Results and docs: GitHub
- Evaluation set: Hugging Face dataset
If you're building for French-speaking Canadians (a Québec clinic, a bilingual help desk, a francophone service outside Québec), start here, tune it on your domain and test it in both languages. Québec data you have the rights to share, an evaluation in your field, or a fix can go back through the GitHub repository.
Where help would be great
The next steps need more GPUs: searching where the bits matter most across a whole model, running the full bilingual test suite on every candidate, and trying training ideas side by side.
If your organisation (a lab, a Canadian compute provider, a company that needs bilingual AI) can provide GPU time or wants to collaborate, get in touch. The results stay open: models on Hugging Face, documentation and evaluations in public.
This work builds on Qwen open models. It is an independent project by Humanjava Enterprises and is not made or endorsed by the Qwen team or Unsloth.



