Can prop firms actually run ONNX-based EAs? (Yes, with caveats.)

Short answer: yes. All major prop firms (FTMO, FundedNext, The5ers, Hola Prime) allow Expert Advisors that load and run ONNX models, because to them it's just a standard EA — the ML inside is invisible. The longer answer involves three caveats that catch people.

Caveat 1: Build 5572 has to be on their MT5 install

Prop firms run MT5 on their own infrastructure (or via a broker partner). If their installation hasn't been updated to Build 5572 or later, the new ONNX_USE_CPU_ONLY flag won't compile, and old ONNX_CUDA_DISABLE flags will work. Most major firms updated within weeks of the January 2026 release.

Verify by attaching an EA with Print(__MT5VERSION__); in OnInit — if the build is below 5572, ask their support.

Caveat 2: The model file has to come with you

Prop firms typically don't give you filesystem access to upload arbitrary files to their MT5 install. If your EA loads the ONNX from MQL5\Files\, you need to get the file there — which they may or may not allow.

The fix: embed the model as a #resource inside the compiled .ex5. The model travels inside the binary — no separate file needed.

The 1 GB resource size limit (raised in Build 5572) covers any retail-scale model. See the OnnxCreate reference.

Caveat 3: No GPU on prop-firm servers (use CPU)

Prop firms run MT5 on Windows VPSes or dedicated servers without NVIDIA GPUs. CUDA inference is unavailable. You must use ONNX_USE_CPU_ONLY.

For most retail-scale models, this is fine: a 100k-parameter LSTM runs in single-digit milliseconds on a modern CPU. For very large models, CPU may be too slow; in that case the prop firm isn't the right home for that specific EA. Use the model only as a filter (called on bar close, not every tick) and you're fine.

What the firms officially allow (and don't)

What every firm prohibits, ONNX or not:

• Latency arbitrage — exploiting slow broker quotes.
• HFT — sub-second scalping at high rates.
• News exploitation — trading on macro releases within seconds.
• Hedging across accounts — opening opposite positions at multiple firms simultaneously.

What every firm allows:

• Standard EAs trading on technical signals, including ML-augmented ones.
• Multi-strategy portfolios as long as risk-management is contained.
• EAs that hold positions overnight (most firms).

The practical workflow

1. Train and export your ONNX model locally or in the cloud.
2. Build the EA with #resource embedding the model.
3. Compile with ONNX_USE_CPU_ONLY in OnnxCreate (prop firms don't have GPUs).
4. Test the full EA on a demo account from the same broker the firm uses.
5. Apply for the prop firm challenge. Pass it with the EA running.
6. Receive funded account. The same .ex5 runs on the funded MT5 install.