North ML

North ML

North ML is an independent artificial intelligence and software engineering team focused on building practical machine-learning systems, developer tools, and experimental language models.

The team works directly on model training, dataset development, evaluation, inference, and software infrastructure. Rather than relying only on third-party AI APIs, North ML experiments with its own models and training pipelines to better understand how data quality, architecture, optimization, and curriculum design shape model behavior.

Current Flagships

Usable daily driver AI models created by North ML. If you want small, gpt-3 class performance on pocket-sized devices, use Aurora/Ember Proelia. Even though more parameters, and less on benchmarks than small models, it delivers frontier-grade performance outside the benchmarks for most edge tasks, and it was trained mostly on a 5060 Ti.

If you want gpt-4.5 class agentic performance, or the ability to research much, use our Qwen-based Qwen3.8-Smol model. It delivers great capability for a 0.8B model, while still being in the range of most bigger models. It can code simple html and stuff. Use this as your creative workhorse.

If you concretely want the best performance, and have enough vram/system memory, use starlight-mini. It has many GGUF's and was trained on Google Colab. It's a ancient model that's still well-trained up to today's standard.

Models in training

ardent-350m (real name undecided, this is a codename for a future model.)

Model Series

• Starlight Mini (one model, discontinued series)
• Wind series (performed well, but it went.)
• Forge series (Possible return near the end of 2026, if I feel like it.)
• Aurora series (Properiatry at first, but any aurora model released not Aurora one is the open-source variant of the new Ember propieriatery series.)
• Ember series (New architecturally fundemental flagship models of the North ML series, starting releases near August 2026.)

Compute Info

• M4 Mac Mini (16 gb variant, Jul 2025)
• M4 MacBook Air (16 gb variant, Dec 2025)
• v6e-16 tpu (Google Tpu research cloud, Feb 27, 2026 - April 31, 2026)
• Rtx Pro 6000 (Google Cloud, Janurary/Feburary - May 2026, then Modal)
• 8x L4 (Trained Aurora One, Modal)
• L40S (Modal)
• Other Modal GPUs.
• 5060 Ti (Jul 2026)

Announcement (May 7, 2026 AD, 3:52 PM)

A new release is expected on May 8, possibly Wind Arc 1.7, Wind Edge 1.7, or something new altogether. Thank you for using our models. I hope you enjoy what’s next.

Announcement (May 10, 2026 AD, 9:59 PM)

Thanks to our friends at Modal for giving us ~$180 in free credit to use B200, L4, and other GPUs.