← Back to Articles
April 15, 2026

The AI Tooling Ecosystem: What's Rising, What's Dying

Tooling

The AI tooling space evolves monthly. Here's my assessment of what's working, what's not, and what's emerging in 2026:

Rising

1. ONNX Runtime

  • Framework-agnostic inference
  • Local + cloud, CPU + GPU + NPU
  • What I use for local inference

    • 2. Vercel AI SDK

  • Simple streaming + tools
  • Works across providers
  • The "default" for web AI apps

    • 3. Ollama

  • Pull and run models locally
  • CLI-first, minimal fuss
  • Best for local experimentation

    • 4. Open Weights Models

  • Qwen, Phi, Llama variants
  • No API dependency
  • Good enough for most tasks

    • Declining

    1. Custom Model Training

    Except for specific domains, fine-tuning isn't worth it. The gap between base models and fine-tuned is shrinking.

    2. Vector Databases for Small Use

  • Postgres has vector support now
  • For most apps, PG or SQLite is enough
  • Dedicated vector DBs are overkill unless at scale

    • 3. Framework-Locked Solutions

  • If it only works with OpenAI, it's a risk
  • Multi-provider support is now baseline

    • What's Emerging

    Agentic Frameworks

    AutoGen, LangGraph, CrewAI—the agent orchestration space is consolidating.

    Edge Deployment

  • WASM-based inference
  • Browser-run models (WebGPU)
  • The "local" extends to client-side

    • Sound and Video

  • Real-time voice integration
  • Video generation (Sora-class)
  • Not just text anymore

    • My Stack

  • ONNX Runtime (inference)
  • Qwen 0.5B (local model)
  • React + Modern.js (frontend)
  • Python backend (flexibility)

    • Simple, replaceable, local-first. That's the philosophy.

    Article 7 of 10 - AI Industry Series