While the giant models get headlines, small ones quietly became very good. In October 2026 Google released EmbeddingGemma 2, an open model small enough to run on a device, and Recursive released Flint, a coding model that runs on a laptop at around 131 tokens per second. Analysts note that a 27-billion-parameter model from Alibaba now competes with models sixty times larger.
This matters in Southeast Asia. Banks, hospitals, government agencies and factories all hold data that cannot leave the country — or the building. Small models turn that rule from a blocker into a design choice: search, classification and first-line assistants can run entirely inside your own walls, with cloud models kept for non-sensitive work.
What it means for your business
- Sort your data by sensitivity first; it decides which model can run where.
- Use small open models for high-volume, simple decisions.
- Weights you have downloaded cannot be taken away; a vendor's promise can.
How MTH can help: We design hybrid AI setups — small models on-premise, frontier models in the cloud — that satisfy UU PDP and sector regulators.
This briefing is original commentary by MTH based on public reporting; figures are as reported by the sources named above and may be revised over time.
