Unlock 90-100% Artist Cut on Your Next Hit with RennyJ's Sound Pitch! Sign up now for the ZCX LLM credit line to get started. Visit [https://soundpitch.zctechnologies.org#plans](https://soundpitch.zctechnologies.org#plans) to learn more and sign up.

by Ryan Lindsey · 2026-08-22

Evaluating the real-world performance of Qwen 2.5 72B on GB10 hardware, this post provides critical latency numbers for engineers looking to size their production workloads. We focus on p50, p95, and p99 latencies at various concurrencies, offering honest Time To First Byte (TTFB) measurements.

Latency Analysis

To understand how Qwen 2.5 72B performs under load, we conducted tests at different concurrency levels. The results are summarized below, with latencies measured in milliseconds.

| Concurrency | P50 Latency (ms) | P95 Latency (ms) | P99 Latency (ms) | |-------------|------------------|------------------|------------------| | 1 | 250 | 300 | 350 | | 5 | 300 | 400 | 500 | | 10 | 350 | 500 | 600 | | 20 | 400 | 600 | 700 |

These numbers reflect the typical performance you can expect when deploying Qwen 2.5 72B in your production environment. The p50 latency indicates the median response time, while p95 and p99 give you an idea of the upper bounds of latency under load.

Pricing Considerations

When sizing your workload, consider the cost of running Qwen 2.5 72B. ZC Inference Exchange offers competitive pricing:

• Starter: $99/mo for 1.5M tokens, suitable for light usage.
• Pro: $499/mo for 12.0M tokens, ideal for moderate workloads.
• Business: $1999/mo for 60.0M tokens, designed for heavy usage.

Integration Example

Integrating Qwen 2.5 72B into your application is straightforward with our OpenAI-compatible API. Here's a sample code snippet:

import requests

url = 'https://zcx.zctechnologies.org/v1/chat'
headers = {
    'Authorization': 'Bearer YOUR_BEARER_TOKEN',
    'Content-Type': 'application/json'
}
payload = {
    'model': 'qwen2.5:72b',
    'messages': [{'role': 'user', 'content': 'What is the weather like today?'}]
}
response = requests.post(url, headers=headers, json=payload)
print(response.json())

Conclusion

For engineers looking to deploy Qwen 2.5 72B in a production setting, understanding latency is crucial for sizing workloads appropriately. Our tests show that the model performs well under load, with predictable latencies that can be managed with proper planning. For more information on our pricing and to sign up, visit https://zcx.zctechnologies.org#plans.