Field notes on building custom AI models: fine-tuning, on-device inference, and the cost maths behind owning your own models.
Google's FunctionGemma (270M) and Gemma 4 E2B (2B) are the smallest credible function-calling models in 2026. They're not general-purpose — they're explicitly designed to be fine-tuned. That's the whole point.
Meta's Llama Stack is the canonical reference architecture for Llama-based agents. Combine it with a fine-tuned Llama 4 derivative and the Swift/Kotlin client SDKs and you get a complete agent stack running entirely on the user's phone.
TypeScript-first mobile builders don't have to use Python agent frameworks. Mastra and the Vercel AI SDK plus a fine-tuned 4B model running on-device through llama.cpp produce a complete agent stack with zero per-token costs.
We benchmarked the three best on-device tool-calling bases of 2026 — Qwen3-4B, Gemma 4 E4B, and Phi-4-Mini — across BFCL v4, real mobile latency, and post-fine-tune accuracy. Each wins a different scenario; here's how to pick.
Deploy custom AI models, no ML expertise required.
Free plan, no card. Paid plans from $25/mo USD.