I’m a C/C++ architect focused on low-latency systems. Last year, I tried building agentic pipelines with SLMs/LLMs and hit the usual wall: latency and orchestration overhead killed real-time edge use cases.

Initial research video: https://www.youtube.com/watch?v=q-rs9VZ1-0I

So I asked: how far can you push specialised logic at <1M parameters with nothing but local CPU?

MicroGPT-C is a from-scratch C99 port of Karpathy’s microgpt.py (https://gist.github.com/karpathy/8627fe009c40f57531cb1836010...). Zero deps, single-header, localised KV cache. Speed was never the goal (Andrej’s nanoGPT and llm.c already showed what’s possible). The real experiment was orchestration.

Organelle Pipeline Architecture (OPA): Agile-style Planners, Workers and Judges talking through tiny structured strings (board=XO_|valid=1,3) parsed by a safety-gated stack VM (3.7–5.8M ops/s). A 64K model needed 181 manual interventions; a 460K model trained on those traces internalised everything and needed zero.

Beyond the research: fully auditable AI, great for education (~3k lines of readable C), rapid prototyping, and embedded.

Personal itch: fraud/risk engines. I want agents that hunt “unknown-unknowns” in a sandbox where every decision is inspectable.

Happy to talk implementation, the 97 tests, 22 benchmarks, or anything else. FAQ: https://github.com/enjector/microgpt-c/blob/main/FAQ.md

Quick try (macOS/Linux/Windows): git clone https://github.com/enjector/microgpt-c && cd microgpt-c mkdir build && cd build && cmake .. -DCMAKE_BUILD_TYPE=Release && cmake --build . ./connect4_demo # 460K params, ~21 min train, 88% win rate vs random

Performance (Apple M2 Max): - 4.2K params names: 685k tok/s train, 110k tok/s infer - 841K Shakespeare char: 28k / 16k tok/s - 510K Shakespeare word: 12.5k / 40k tok/s

Full leaderboard (11 games), market-regime experiment (57% holdout = 2.8× baseline), and the book PDF: https://github.com/enjector/microgpt-c/blob/main/docs/book/M... GitHub: https://github.com/enjector/microgpt-c