Command Code Claims Fastest DeepSeek Inference at 197 Tokens/s in Head-to-Head Benchmark

Ahmad Awais 1 hour ago

Key Info

Command Code's operator reports that its route for DeepSeek V4.1 Flash achieved the highest median effective throughput in a real-call comparison, at 197 tokens per second—2.6x faster than the slowest tested route.

Highlights

Measured 2,679 real calls with 2,000+ output tokens each across four providers/routes, reporting median effective throughput: Command Code 197 tok/s, OpenCode Go 186 tok/s, ClinePass 122 tok/s, Ollama Cloud 75 tok/s.
The results are wall-time measurements of full requests, not provider-reported raw decode speeds, so they reflect end-to-end latency differences.
The post frames this as validation of Command Code's "inference-native" focus on harness and inference engineering, positioning the service as faster and cheaper for the same model.

Original post

Command Code is fastest DeepSeek inference in the world. Better Faster Cheaper All the work we put into tool repairs have and harness + inference engineering is paying dividends now. Our mission is to be inference-native. Let's go, GOAT 🐐

Quoted post

DeepSeek V4.1 Flash is not one speed. Same model. Same Hermes setup. Very different throughput depending on the route. I pulled 2,679 real calls with 2,000+ output tokens across four providers. Median effective throughput: CommandCode GOAT: 197 tok/s OpenCode Go: 186 tok/s ClinePass: 122 tok/s Ollama Cloud: 75 tok/s That’s a 2.6x difference between the fastest and slowest route for the same model, same harness. These are full request wall-time measurements, not provider-reported raw decode speeds. I filtered to long responses so prefill matters much less. Provider choice matters way more than I expected.

#DeepSeek#inference speed#Command Code

View original on X