Français English

Language models, sized to fit.

Aulvero trains compact bilingual models for French and English, from fifty million to nine hundred million parameters. They're built to run on the hardware people already own.

Wren50MFinch85MRobin125MLark160MKestrel250MOsprey350MHeron500MCrane700MCondor900M
Fig. 1Every model on a logarithmic scale. Each step is roughly 1.4× the one before it.Select a model to open its page.

Three tiers, chosen by where they run.

Most people start with the hardware they have, not a parameter count. So that's how we group the models.

Tiny

50M – 125M

Phones, browsers, embedded devices

Small enough to run inside a web page or on low-power devices. Fast and inexpensive, and strong on narrow tasks once fine-tuned.

Small

160M – 350M

Phones and laptops

The everyday middle. Runs comfortably on a phone or any laptop CPU, with enough capacity for general instruction following.

Medium

500M – 900M

Laptops and single GPUs

Our most capable models. Still small enough for one consumer GPU or a recent laptop, with noticeably stronger reasoning and writing.

Every size, more than one way.

Each model ships as a Base checkpoint and an Instruct checkpoint fine-tuned from it. A Code variant comes later.

Base Pre-trained

The raw pre-trained model. It continues text rather than following instructions. Use it as the starting point for your own fine-tuning.

Pre-training on French, English and code

Instruct Follows instructions

Fine-tuned from Base to follow instructions and hold a conversation in French or English. Most people should start here.

Supervised fine-tuning of Base

Code Coming later

Base models further pre-trained on a code-heavy mix with fill-in-the-middle, for completion and repository understanding.

Continued pre-training on code

How we train.

One architecture and one codebase cover every size. The same four steps take each model from raw text to a download.

  1. 01

    Data

    French web text from FineWeb-2, English educational text from FineWeb-Edu, and permissively licensed source code. Everything is cleaned, filtered and de-duplicated.

    One 49,152-token tokenizer for both languages and code
  2. 02

    Pre-training

    Each size trains from scratch on roughly a hundred tokens per parameter. A warmup-stable-decay schedule lets us extend promising runs without starting over.

    Produces the Base checkpoints
  3. 03

    Instruction tuning

    Base models are fine-tuned on French and English conversations, and learn only from the assistant's side of each exchange.

    Produces the Instruct checkpoints
  4. 04

    Release

    Weights are exported in a standard Hugging Face format, so they load directly in Transformers, vLLM, llama.cpp and Ollama.

    No custom code needed to run a model
aulvero-train
# one command per stage, the same for every size
$ make prepare
$ make pretrain MODEL=kestrel-250m GPUS=8
$ make finetune MODEL=kestrel-250m GPUS=8
$ make export   MODEL=kestrel-250m

Reading a model name.

The names follow birds, smallest to largest. The rest of the name says exactly what you're downloading.

KestrelNameone per size
250MSizeparameters
InstructVariantBase or Instruct
v1Versiongeneration
huggingface.co/Aulvero/Kestrel-250M-Instruct-v1

Follow the lineup as it trains.