I’m sorry, but I can’t give you a definitive recommendation without the missing pieces of information.  
Here’s a quick framework you can use once you have the full picture:

| Decision factor | What you need | Why it matters |
|-----------------|---------------|----------------|
| **Accuracy** | You already have it | Determines how often the model gives the right answer. |
| **Latency** | You already have it | Determines how quickly the model can respond in production. |
| **Memory footprint** | Unknown | Affects the size of the VM/containers you’ll need and can influence cost. |
| **Failure rate / reliability** | Unknown | Determines how often the model will crash or return unusable results. |
| **Operational cost** | Derived from memory, latency, and failure rate | Directly impacts your budget. |
| **Business impact** | Depends on accuracy and latency | Determines how much value the model brings. |

### What to do next

1. **Collect the missing data**  
   * Run a small‑scale pilot for each model to measure peak memory usage and failure rate under realistic load.  
   * If you can’t run a pilot, look for published benchmarks or vendor data that include those metrics.

2. **Quantify the trade‑offs**  
   * Create a weighted scorecard (e.g., accuracy × 0.4 + latency × 0.3 + memory × 0.2 + failure × 0.1).  
   * Adjust the weights to reflect your business priorities (e.g., if latency is critical, give it more weight).

3. **Risk assessment**  
   * If one model has a significantly higher failure rate, consider whether you can tolerate that risk or if you need a more robust solution.

4. **Make the call**  
   * Once you have all four metrics, compare the weighted scores.  
   * If the scores are close, look at secondary factors such as vendor support, ease of deployment, or future scalability.

Until you have the memory and failure‑rate data, I can’t say which model is the better choice. Gathering those metrics will let you make a data‑driven decision.