Small models that check their work.
An independent lab building compact open models, and the stack that lets them route, search, verify and act.
North ML · October 2026 · north-ml.space

Grounded AI is priced for the few.
Frontier APIs
$2–$50
per million output tokens. Closed weights, cloud only, and your data leaves your machine.
Local small models
Cheap, private
but stale knowledge, made-up facts and no tools. Fluent, not reliable.
The gap isn't raw parameters. Small models can't check their work.
Make the system smart, not just the weights.
- RouteA 1,538-parameter router decides whether to search and whether to think.
- SearchLive web and Wikipedia, reranked with BM25 and embeddings, passages picked per question.
- VerifyFiber-1 Pro checks every claim and number against the sources before answering.
- ActPython, maths, time, weather and geocoding tools ship with the runtime.
The model is one component. Routing, search, state and tools are the product.
The Flare-1 lineup.
| Model | Parameters | Size (8-bit) | What it adds |
|---|---|---|---|
| Ink-1 | ~725M | 755 MB | Starlight Small backbone with the Kestrel base |
| Fiber-1 | ~1.63B | 1.76 GB | Starlight Pro backbone with the Kestrel base |
| Fiber-1 Pro | ~1.63B | 1.76 GB | Online GRPO plus a neural fact-verification pass |
| Flare-Preview | 1M-class | – | Architecture preview and a published negative result |
All Apache-2.0 on Hugging Face.
One file, one Mac mini.
- Starlight backbone 24 layers: three gated DeltaNet, then one GQA attention, repeated.
- Fiber trunk a 158M transformer that routes, drafts tokens and holds associative memory.
- AgentRegisters memory slots whose state survives across calls.
- Era lossless snapshots and rollback of every state, for speculative decoding.
- Kestrel runtime MLX, streaming each stage as it happens.

Measured, then published.
| Benchmark | Ink-1 | Fiber-1 |
|---|---|---|
| ARC-Challenge (1,165 q) | 63.6% | 73.5% |
| MMLU (280 q subset) | 39.3% | 48.9% |
| HumanEval (164) | 18.3% | 37.2% |
| GSM8K (300 q) | 40.3% | unstable, flagged |
19/20
Fiber-1 Pro on a held-out reliability eval (Fiber-1 and Ink-1: 17/20), with fictional entities and offline sources.
We publish failures too: Flare-1M showed no significant advantage over a matched GPT baseline, and we said so. English only; no claim of frontier parity.
Thirteen months on Hugging Face.
All-time downloads of the models released in each month, North-ML1 and arthu1 accounts. Source: Hugging Face API, 21 October 2026.
Shipping, every month.
- First uploadssnowy, Neura-1, then astrocoder-star
- starlight-minifirst North-ML1 release
- North Star familytrained from scratch on one Mac mini
- Wind-Edge 1.6 GGUF3.8k downloads; starlight-mini GGUF 4.3k
- Willow, Forge-1, Curriculum11 releases in a month
- Aurora One Mini3.2k downloads, 6 likes
- Qwen3.8-Smol-uo, Aurora Proelia11.4k downloads for Smol-uo
- Starlight, then Flare-1Ink-1, Fiber-1, Fiber-1 Pro
What people can use today.
Parallax
A research workspace: plans, searches the live web, cites sources. Ink 1, Fiber 1 Lite, Fiber 1 Pro and Mix.
API
OpenAI-compatible and streaming, with a dashboard, keys and usage logs. One model, north-auto.
Solace
An autonomous coding app that runs its own code and repairs it until tests pass.
Open weights
Every model Apache-2.0, many as GGUF. Download, run, fine-tune.
Accounts are approved by hand during the beta.
What's next.
Known limits, published
- English only
- Hallucination is still possible
- The router is a linear probe
- One Mac mini serves everything
Plans
- More inference capacity beyond a single machine
- Longer-context training to test the Flare mechanisms properly
- A stronger router, then more languages
Everything we learn, including what fails, gets published.
Run it on your own desk.
Weights on Hugging Face. Papers on the site. An API and a research workspace for anyone we let in.
- Sitenorth-ml.space
- Weightshuggingface.co/North-ML1
- Researchnorth-ml.space/research
- Try the APInorth-ml.space/try