Research · North ML
Small models,
measured in public.
Every paper here ships next to weights you can download and check. Results and limitations sit in the same size type, and the experiments that failed are written up with the ones that worked.
- Papers
- 3
- Models documented
- 7
- Failures written up
- 3
- Lead researcher
- Arthur Y.
Latest · October 2026 · Technical report
Flare-1A merged hybrid architecture for grounded, tool-using small language models
The architecture behind Ink-1, Fiber-1 and Fiber-1 Pro: a gated DeltaNet/GQA backbone, a second Fiber trunk, AgentRegisters, DeltaMemory, a router head and the Era state manager, with measured results and a published negative result.
- ARC-Challenge, Fiber-1
- 73.5%
- HumanEval, Fiber-1
- 37.2%
- Held-out reliability, Fiber-1 Pro
- 19/20
- Parameters, Fiber-1
- 1.63B
All papers
Flare-1 A merged hybrid architecture for grounded, tool-using small language models
The architecture behind Ink-1, Fiber-1 and Fiber-1 Pro: a gated DeltaNet/GQA backbone, a second Fiber trunk, AgentRegisters, DeltaMemory, a router head and the Era state manager, with measured results and a published negative result.
Wind-Edge-1.6 Compact distillation for edge assistant inference
Depth pruning from 28 to 18 blocks, a healing pass and multi-stage distillation into a 0.44B assistant, and what went wrong on the first instruct attempt.
North Star Building Christian-grounded language models from scratch on consumer hardware
North Air 1, North Star 1 and Wind Arc 1.5, trained from scratch on one M4 Mac mini: a layer-duplication growth strategy and teacher-logit caching for distillation on 16 GB.
No paper matches that.
Lineage
Each paper documents a model family. Read left to right to see how the lab got here.
- North Star
North Air 1North Star 1Wind Arc 1.5
- Wind-Edge-1.6
Wind-Edge-1.6-Instruct
- Flare-1
Ink-1Fiber-1Fiber-1 Pro
How we publish
Weights ship with the paper
Every claim points to an artifact you can download and run on your own machine.
Failures in the same type
Negative results get their own section, their own numbers and a link from the top of the page.
Corrections are welcome
Send replication reports to beta@north-ml.space. Arthur Y. signs off on every paper and reads every report.
More: how the models are built · pitch deck · weights on Hugging Face