Cognitive cores

targeting 1.5 – 2 B params

On-device reasoning. Laptop and phone budgets.

spec

• INT8 weights, FP16 activations on high-sensitivity layers.
• Context window: 8 K tokens initial, scaling pending memory profile.
• Distilled from a frontier teacher. No fine-tune fork from a public base.

target deploy

WebGPU on laptops, Apple Neural Engine on phone-class silicon, commodity ARM via llama.cpp.

eval plan

A frozen suite covering reasoning, instruction following, and known refusal patterns. Reported as the geomean against a stated FP32 teacher baseline; per-task numbers in the model card.

targeting ~3 B params

Single GPU production traffic.

spec

• Mixed-precision: INT8 weights, FP16 activations at attention layers.
• Context window: 32 K tokens. Sliding-window attention beyond.
• Calibrated against the deployment workload, not a public set.

target deploy

Single consumer-class GPU (RTX 4070 and up), or one A10G / L4 in production.

eval plan

Same frozen suite as Edge core plus a workload-specific eval drawn from the deployment domain. Sensitivity per layer published with the model card.

targeting ~7 B params

Distilled from frontier teachers, longer contexts.

spec

• INT8 weights, FP16 activations. INT4 weights as an opt-in compile flag.
• Context window: 128 K tokens, retrieval-augmented by default.
• Distilled from a frontier teacher with curriculum from the task graph.

target deploy

Single A100, H100, or equivalent. Multi-replica scale-out via standard inference frameworks.

eval plan

Standard suite plus long-context retrieval evals (NIAH, etc.). Reproducibility hash and seed published per release.