Qwen 3.7 max has dropped. It’s available through Singapore and it’s 🤑💵🤑!
I have to say, it’s an incredible model, easily comparable to Claude Opus capability, far exceeds the best open-weight models I can run.
My challenge to it was to refactor a 1200 lines of python AI slop into a proper classes. It did it and did so quickly. I’m pretty happy with the code. Qwen3.6-27B and Qwen3.6-35B-A3B cannot refactor this code (at least not with a 10 minute timeout on any single instruction).
Summary
Created HarnessRunner class with clean separation of concerns:
### Initialization (__init__)
- _resolve_config(args) — extracts all CLI arguments into instance attributes
- _init_model() — loads the indicator detection model
- _init_hud() — initializes HUD detector with confidence thresholds
- _init_game_state() — sets up game state tracking
- _init_pad() — creates virtual gamepad
- _init_goal(args) — initializes waypoint goal controller
- _init_grabber() — sets up screen capture
- _init_pid() — configures PID controller with nav/compass profiles
- _init_log() — opens JSON log file
- _init_loop_state() — initializes frame-to-frame state
### Runtime methods
- run() — main loop wrapper with exception handling
- _tick() — single frame processing
- _capture_and_normalize() — grab and resize frame
- _check_hotkeys() — handle keyboard input
- _perceive(t) — detect nav marker or compass indicator
- _perceive_nav() / _perceive_compass() — perception sub-routines
- _compute_control(dt, active) — PID calculation
- _apply_control(active) — smooth and send commands
- _log_frame() — write JSON log entry
- _update_fps() — rolling FPS average
- _status_string() — format status label
- _update_overlay() — render overlay window
- _print_status() — console output
- _cleanup() — release resources
### Improvements
- All args.X references resolved once in __init__, stored as self.X
- No self._args or args. in runtime methods
- Dead code removed: do_calibrate(), draw_fullscreen_overlay(), CALIB_PATH, nav_arrived
- PID controller extracted to pid_controller.py (168 lines)
- Per-frame state uses self._* prefix for clarity
Files: 1307 lines (harness-compass.py) + 168 lines (pid_controller.py)
Alibaba cloud for coding you need a token subscription – I went with $30 and that 10 minute or so refactor of 1200 lines of code cost 20% of my monthly allowance!
The models coming of China are as good as those coming out of the US – they also cost about as much!
#AI #Qwen #AISoftwareEngineering #softwareengineering
