Edge AI means on-device LLMs keep data local and skip API bills. Q4 3B–7B models already produce tokens on machines like iPhone 15 Pro and MacBook M3. Ollama is the fastest entry; Apple Silicon uses MLX, everything else uses llama.cpp. Fit for offline drafts and privacy-sensitive queries, not as a drop-in cloud-model replacement.

Early 2026 Edge AI is already usable.

iPhone 15 Pro + 3B Q4 model, daily simple tasks sufficient. Mac Studio can run 13B model, most scenarios comparable to cloud API.

Codestral is the current reason to keep a specialized code model in the stack: fill-in-the-middle at IDE latency, not another chat endpoint. General models in the GPT-5.6 family, Claude, and Gemini still own planning, tools, and mixed context. No local bake-off, so no scoreboard. The only numbers below are vendor claims with a live official URL.

AI Agents that autonomously complete multi-step tasks. No step-by-step human steering—give a goal, it plans, executes, verifies.

Public writeups often name Devin, Manus, and Cursor Agent in one breath. No controlled bake-off of the three. No scores.

One independently built Go tree, one go.mod. The gate is gofmt, go test ./..., and go vet ./.... The rest of this note is the boundaries a production tree actually hits: modules, errors, concurrency, tests. Effective Go is not recopied.

OpenAI vision is a shipping API, not a 2024 demo. So are Gemini and Claude vision. What matured is the endpoint. What did not mature is an unsourced leaderboard. No bake-off was run for this post, so no accuracy column.

OpenAI drops GPT-5.4.

Official positioning: “our most capable and efficient frontier model for professional work.”