Field notes
Small models that check their work.
North ML builds compact open models that route, search, verify and act. Apache-2.0, and small enough to run on a 16 GB Mac mini.
The lab
An independent lab, with a short list of beliefs.
Open weights first
Every model ships under Apache-2.0 on Hugging Face, many with GGUF builds. If we publish a number, you can download the thing that produced it and check it yourself.
Efficiency over scale
Strong results per parameter and per dollar, not frontier-size clusters. The whole Flare-1 lineup is served from one Mac mini.
Systems thinking
The model is one component. Routing, search, state and tools are the product. A small model that can look something up and check its claims closes most of the gap that raw parameters would.
Everything published
Including failures. When our 1M-parameter experiment showed no advantage over its baseline, we wrote it up anyway. Limitations sit next to results, in the same size type.
Kestrel
One question, four passes.
The runtime holds still, looks, and only then commits. The API streams each stage as it happens, so you can watch it hover.
p_search 0.91, p_think 0.08. Decision: search.
Query “North ML Flare-1 release”. Six candidates, two kept after rerank.
“October 16 to 17, 2026” matches. “1M users” is removed.
A short, cited answer, with the two sources it kept.
Inside Flare-1
One merged file per tier.
Backbone, trunk, router, bridge, tokenizer and runtime metadata in a single safetensors file, deployable on a 16 GB Mac mini.
Starlight backbone
24 layers: three gated DeltaNet layers, then one full GQA attention layer, repeated. Pruned and distilled from a larger Qwen3.5-style teacher.
AgentRegisters
Persistent memory slots after each full-attention layer. State survives across calls, snapshots and rollbacks. 295k parameters in Ink, 590k in Fiber.
Fiber trunk
A complete 24-layer transformer that shares the backbone’s embeddings. It feeds the router, drafts tokens for speculative decoding, and holds associative memory.
Router head
A 1,538-parameter linear probe that decides whether to search and whether to think before answering. It gates the retrieval pipeline. Overrides: code always thinks; small talk never searches.
Bridge
A learned linear projection from Starlight’s hidden states into the Fiber trunk’s width, so the two halves can read each other. 2048 to 768 in Fiber-1 (3.15M parameters); 1024 to 768 in Ink-1 (1.57M).
Era
A lossless state manager: snapshot and roll back the KV cache, DeltaNet states, registers and memory. Exact rollback is what makes speculative decoding safe.
Measured, then published
What it scores, and what we don’t claim.
| Benchmark | Ink-1 · 725M | Fiber-1 · 1.63B |
|---|---|---|
| ARC-Challenge 1,165 questions | 63.6% | 73.5% |
| MMLU 280-question subset | 39.3% | 48.9% |
| HumanEval 164 | 18.3% | 37.2% |
| GSM8K 300 questions | 40.3% | unstable, flagged |
Held-out reliability eval (chat, identity, grounding, follow-up, pushback, reasoning, code): Fiber-1 Pro 19 of 20, Fiber-1 17 of 20, Ink-1 17 of 20.
A result that didn’t work. Flare-1M (965,908 parameters) against a 981,312-parameter GPT baseline: no significant advantage, recall 0.146 against 0.104. Inconclusive, because the task outgrew the 384-token training window. It’s in the report.
What you can use
Four ways in.
Aspect
Research projects that plan a question, read sources for each part, cross-check, and write a cited paper, plus Quick Chat for short sourced answers. All on North ML’s served models. Open Aspect.
API
OpenAI-compatible streaming chat completions with a console, keys and usage logs. north-auto routes each request to Ink-1 or Fiber-1. Read the docs.
Solace
A desktop app for research and code that signs in through your browser. Download it with an approved account.
Open weights
Ink-1, Fiber-1 and Fiber-1 Pro as safetensors, many earlier models as GGUF. Download, run, fine-tune. Hugging Face.