Skip to main content

You walk the city. The AI calls the plays.

A 0.5B-parameter Qwen 2.5 PPO model is your Operator. Every turn it decides what to do — negotiate, scout, deploy, run the op, secure extraction — and drops a glowing waypoint on Neon Meridian's map. Your job: walk there with W A S D and E to execute. The mission only succeeds if the two of you cooperate inside the env's real reward function.

Pick a mission

    The trained checkpoint is partially trained, so raw mode often picks moves the env will penalize (e.g. run_operation before deploy). Hybrid keeps the LLM in the driver seat but rejects actions that would burn turns; the dashboard shows every override. Use Heuristic if you just want a clean win.

    How the heist works

    1. Comms. The Operator (the trained LLM) speaks each turn. Its raw text and parsed action are visible.
    2. Waypoint. A neon ring + arrow appears on the relevant district. That's where you go.
    3. Walk. Use W A S D or ↑←↓→. Vex glides through the iso city.
    4. Execute. When in range, press E. The action commits to the env, VFX fire, the world reacts.
    5. Adapt. The Operator sees the new state and picks the next move. Repeat for up to 12 turns.
    6. Extract. Hit five mission gates and secure the extraction. Don't get timed out.

    Districts of Neon Meridian

    • Docklands — Iron Vultures' turf. Fast, loud.
    • Data Spire — Civic Shield watches every channel.
    • Undergrid — Ghostwire trades intel in the dark.
    • Citadel Gate — Black Orchid moves through influence.

    Accessibility

    Keyboard-first. No timed reflex tests. Reduced motion respected. A spoken/screen-reader summary updates each turn. High-contrast neon palette throughout.

    Neon Syndicate is a partially observable, long-horizon strategy environment for OpenEnv. The Operator is a 0.5B Qwen 2.5 fine-tuned with TRL PPO; the dashboard honestly flags when its output can't be parsed and the heuristic takes over.