Parallel attempts
Four reasoning styles, deterministic seeds, bounded workers, and one shared deadline.
Open-source Python package
Several ways to solve it. One answer to return.
An AIMO3-style math reasoning agent for vLLM and OpenAI-compatible models. It runs parallel solutions, checks calculations in Python, and stops when the answers agree.
The useful work happens around the model.
A model can write a promising solution and still miss the integer, repeat the same mistake eight times, or spend the full budget on one stuck request. I pulled that surrounding logic into a package that can be tested without loading a model.
Python checks
Structured tool calls run in a separate local process with time and output limits.
Stable consensus
Boxed integers are counted first. Token entropy only settles equal-vote cases.
Run the control loop in a minute.
The demo uses fake attempts, so it works on a laptop and does not download model weights.
git clone https://github.com/xufenghe/aimo3-agentic-inference.git
cd aimo3-agentic-inference
python -m venv .venv
source .venv/bin/activate
python -m pip install -e .
aimo3 demoproblem → diverse attempts → Python checks → boxed answers → vote → result
Made for local model servers.
The included client uses Chat Completions and works with vLLM-style endpoints. The documented example uses gpt-oss-20b, but the model name is configurable.
Questions that come up
Is this an official AIMO project?
No. It is an independent clean-room implementation inspired by the AI Mathematical Olympiad ecosystem.
Does it need a GPU?
The package, tests, and demo do not. Real inference needs a local or remote model endpoint.
Is Python execution sandboxed?
Not against hostile code. The local runner is for trusted experiments; untrusted code belongs in a container or microVM.
Does it include a benchmark result?
Not yet. The evaluator is included so results can be published with the exact model and settings later.