Making aime™ smarter on smaller models — a system-wide move that cuts LLM cost by 60–80% without losing quality

aime today described, at a principles level, a platform shift that runs the full aime agent stack on small-class models rather than frontier-scale LLMs. The result, targeted across the platform, is a 60–80% reduction in LLM cost and prompt complexity, faster response times, and the freedom to run aime on cheaper infrastructure — with no loss in classroom quality.

The core idea

Most of what aime's agents do today does not actually need a frontier model. A great deal of agent work is structure, lookup and arithmetic. aime was paying frontier-model rates for tasks that belong in code. The shift is simple: only ask a model to do what only a model can do — generate human-quality content — and move counting, sorting, numbering, computing and formatting into deterministic code.

“We were paying frontier prices for arithmetic. The breakthrough is not a bigger model — it is recognising which work belongs in a model at all, and moving everything else into deterministic code.”

The posture

• Only the model does model-grade work — every other step is deterministic.
• Reasoning is selective — applied where it materially improves output, suppressed where it does not.
• Schemas and prompts are shaped for compact models, not assumed to inherit frontier-model tolerances.
• Validation is treated as a first-class layer, not an afterthought.

The agent-by-agent implementation — including specific prompt restructures, schema decisions, runtime choices and the internal mitigations that take small-model error rates down to production-grade levels — is documented in aime's internal engineering library and shared with partners under NDA.

“We didn't just shrink the prompts. We changed who does the work. The model writes the human-quality content. The code does the counting, the lookups and the structure. That is the whole game with small models.”

Impact at a glance

• 60–80% reduction in prompt complexity and LLM token consumption across the platform.
• Same output quality — measured against aime's existing validators and classroom-readiness checks.
• Faster response times and dramatically lower per-lesson cost.
• Freedom to deploy aime on cheaper infrastructure, including offline-capable environments.

“Personalised education at population scale only works if the unit economics work. Making aime smarter on smaller models is the unlock. A different lesson pack for a teacher in rural India and a different one again for a teacher in the UK — both for pennies, both classroom-ready.”

How aime gets there

The rollout is a one-week, three-phase programme. Days 1–2 ship the easy wins (Kern™ migration, non-thinking-mode swaps, lookup tables). Days 3–4 land the architectural moves (split passes, flatter schemas, pre-fill). Days 5–7 deliver the Lesson Designer breakdown into per-slide-type micro-agents. The work is sequenced inside aime's own aime Loom™ Workflow Engine™ for deterministic, replayable execution.

Availability

The Smarter-on-Smaller-Models programme is in active rollout across the aime platform from May 29, 2026, beginning with Curriculum Generation and extending agent-by-agent across aimeCLOUD™ over the following week.

About aime

aime builds the operating system for educational intelligence — the foundational infrastructure layer that future education systems will run on. aime's stack combines structured curriculum knowledge, pedagogy-aware reasoning, compact education-tuned models, agentic orchestration and offline-capable deployment, and is designed for ministries of education, universities and national school systems.