Ink — North ML Flare-1 (~725M)

The smallest Flare-1 tier: Qwen3.5-derived gated DeltaNet + GQA hybrid backbone with AgentRegisters — persistent learned state slots for agent-style memory. This is the full-precision training checkpoint (HF transformers format). The production 8-bit Kestrel build with the learned router, Fiber memory trunk, and search pipeline is at North-ML1/Ink-1.

Benchmarks (measured, same harness — Modal L4, greedy, Nov-2025)

Model	ARC-C	MMLU	GSM8K	HumanEval
ink (725M)	63.8%	32.0%	30.0%	37.5%
fiber-lite (1.6B)	72.9%	38.0%	16.0%	37.5%
fiber-pro (1.6B)	80.9%	40.0%	14.0%	48.4%
Qwen3-0.6B (raw)	49.2%	29.0%	4.0%	4.7%
Qwen3-1.7B (raw)	72.9%	39.5%	43.0%	68.8%
SmolLM3-3B (raw)	77.9%	36.0%	51.0%	71.9%

Subsets: ARC-C n=199, MMLU n=200, GSM8K n=100, HumanEval n=64. Honest read: strongest size-adjusted scores on knowledge MC in this lineup; trails the newer 1.7B-3B models on math and code.

Specs

Parameters	~725M (565M backbone + 158M Fiber trunk + registers + bridge)
Layers	24
Hidden	1024
Attention	GQA 8 heads / 2 KV, DeltaNet:GQA = 3:1
Context	262,144 tokens
Vocab	65,536

Interactive chat

pip install torch transformers safetensors huggingface_hub
python3 chat.py --model North-ML1/ink

chat.py (in this repo) auto-downloads the weights, attaches the AgentRegisters, and starts a REPL with optional two-phase thinking.

Links

• Architecture report: Flare-1 paper
• Runtime artifacts: Ink-1 · Fiber · Fiber-pro