Performance
LW AI is designed for high throughput and low latency. This page provides performance benchmarks and optimization guidance.
Latency Benchmarks
Typical response times measured from the Hillsboro and Singapore regions:
| Model | Time to First Token (TTFT) | Tokens/sec (Output) |
|---|---|---|
| deepseek/deepseek-v4-flash | ~200ms | ~80 tok/s |
| deepseek/deepseek-v4-pro | ~500ms | ~40 tok/s |
| z-ai/glm-5.1 | ~400ms | ~50 tok/s |
| z-ai/glm-5.2 | ~600ms | ~35 tok/s |
| linkwo/fusion | ~300ms | ~60 tok/s |
| astra | ~400ms | ~45 tok/s |
Values are approximate and vary based on input length, output length, and current load. Measured with short prompts (~100 tokens input).
Regional Latency
| Client Location → API Region | Estimated Latency |
|---|---|
| Singapore → openapi.linkwo.ai | ~15ms |
| Tokyo → openapi.linkwo.ai | ~50ms |
| US West → openapi.linkwo.ai | ~15ms |
| US East → openapi.linkwo.ai | ~60ms |
| Europe → openapi.linkwo.ai | ~200ms |
| South America → openapi.linkwo.ai | ~120ms |
Throughput
| Tier | Max Concurrent Requests | Estimated Max TPM |
|---|---|---|
| Free | 2 | 40,000 |
| Standard | 10 | 200,000 |
| Pro | 30 | 1,000,000 |
| Enterprise | Custom | Custom |
Optimization Tips
Reduce Latency
- Choose the right region — Use
openapi.linkwo.ai(routes to the nearest region automatically) - Use Flash models — DeepSeek V4 Flash is 2-3x faster than Pro
- Enable streaming — Receive tokens as they're generated instead of waiting for the full response
- Shorten prompts — Less input tokens means faster processing
Maximize Throughput
- Batch requests — Use the Batch API for non-real-time workloads
- Connection pooling — Reuse HTTP connections across requests
- Concurrent requests — Send multiple requests in parallel (within rate limits)
- Cache results — Avoid duplicate API calls for identical inputs
Reduce Token Usage
- Set
max_tokens— Limit output length when you don't need long responses - Use concise prompts — Be specific and avoid unnecessary context
- Leverage system prompts — Set behavior once instead of repeating in every message
- Use structured output — Request JSON directly to avoid verbose formatting
Monitoring Your Usage
Contact support@linkwo.com to get performance metrics for your account:
- Request latency percentiles (p50, p90, p99)
- Token usage over time
- Error rates by model
- Rate limit utilization
Related
- Rate Limits — API rate limiting
- Routing & Fallback — Automatic fallback
- Service Status — Real-time status