Start typing to search.
Ranks providers by their recent typical output speed in tokens per second. Providers without enough recent data remain visible after ranked rows.
Weights & Biases currently ranks first at 113.0 tok/s via wandb/bf16.
| Rank | Provider / route | Output speed | Speed measurements | Endpoint details |
|---|---|---|---|---|
| #1 | Provider / route Weights & Biases wandb/bf16 | Output speed113.0 tok/s | Speed measurements
| Endpoint details
|
| #2 | Provider / route Groq groq | Output speed100.0 tok/s | Speed measurements
| Endpoint details
|
| #3 | Provider / route NovitaAI novita/fp8 | Output speed67.0 tok/s | Speed measurements
| Endpoint details
|
| #4 | Provider / route DeepInfra deepinfra/fp8 | Output speed34.0 tok/s | Speed measurements
| Endpoint details
|
| #5 | Provider / route DeepInfra deepinfra/bf16 | Output speed22.0 tok/s | Speed measurements
| Endpoint details
|
| #6 | Provider / route Cloudflare cloudflare/fp8 | Output speed17.0 tok/s | Speed measurements
| Endpoint details
|
Endpoint data fetched .
Llama 3.1 8B Instruct endpoint guide
6 of 6 Llama 3.1 8B Instruct endpoints currently have the published data required for this ranking. Weights & Biases leads at 113.0 tok/s via wandb/bf16. The table keeps unranked routes visible so missing measurements do not look like missing provider availability.
The output-speed ranking compares recent median generated tokens per second. It favors routes that complete long generations quickly, while first-token latency remains a separate measure of how responsive the endpoint feels at the beginning.
Groq currently ranks second at 100.0 tok/s. Compare that gap with input and output price, response-start time, output speed, 30-minute uptime, context length, and the exact route before deciding whether first place is meaningful for your workload.
Llama 3.1 8B Instruct has 6 published provider options, and performance data is matched to each exact OpenRouter routing tag. Different quantization, context, regional deployment, or provider configuration can change price and behavior even when the underlying model name is identical.
