Performance

Time to first token (TTFT) and output speed for each model and completion window, measured on live Sail traffic.

Model	Window	TTFT (median)	TTFT (p95)	Output speed
Kimi K3 `moonshotai/Kimi-K3`	Default (ASAP)	1.3 s	19 s	120 tok/s
	Balanced	2.1 s	6.9 s	108 tok/s
	Flex	1.6 s	5.7 s	137 tok/s
GLM-5.3 `zai-org/GLM-5.3`	Default (ASAP)	1.5 s	58 s	65 tok/s
	Balanced	1.9 s	1.1 min	188 tok/s
	Flex	1.9 s	7.5 s	261 tok/s
GLM-5.3-Flash `zai-org/GLM-5.3-Flash`	Default (ASAP)	1.7 s	1.9 min	25 tok/s
	Balanced	3.8 min	14.5 min	19 tok/s
	Flex	4.4 s	2.4 h	23 tok/s
DeepSeek V4.1 Flash `deepseek-ai/DeepSeek-V4.1-Flash`	Default (ASAP)	3.6 s	31 s	44 tok/s
	Balanced	2.9 s	54 s	55 tok/s
	Flex
DeepSeek V4 Pro 0813 `deepseek-ai/DeepSeek-V4-Pro-0813`	Default (ASAP)	1.6 s	4.2 s	21 tok/s
	Balanced
	Flex			27 tok/s
DeepSeek V4 Flash 0731 `deepseek-ai/DeepSeek-V4-Flash-0731`	Default (ASAP)	1.5 s	6.1 s	82 tok/s
	Balanced	2.2 s	11 s	54 tok/s
	Flex	3.0 s	7.9 s	48 tok/s
Kimi-K2.6 `moonshotai/Kimi-K2.6`	Default (ASAP)	1.1 s	3.0 s	41 tok/s
	Balanced
	Flex
Gemma 4 31B IT `google/gemma-4-31B-it`	Default (ASAP)	0.6 s	6.1 s	67 tok/s
	Balanced	10 s	1.9 min	33 tok/s
	Flex	1.5 s	1.9 min	32 tok/s
Gemma 4 31B IT (NVFP4) `nvidia/Gemma-4-31B-IT-NVFP4`	Default (ASAP)	1.0 s	2.8 s	27 tok/s
	Balanced	0.9 s	2.5 s	34 tok/s
	Flex
Gemma 4 12B IT `google/gemma-4-12B-it`	Default (ASAP)	0.5 s	1.0 s	47 tok/s
	Balanced
	Flex
gpt-oss-120b `openai/gpt-oss-120b`	Default (ASAP)	0.6 s	1.1 s	203 tok/s
Qwen3.6 35B A3B `Qwen/Qwen3.6-35B-A3B`	Flex	0.6 s	1.5 min	47 tok/s

Notes

Time to first token grows with the length of your input, because the model processes the whole prompt before it produces the first token.
TTFT for balanced and flex includes time spent waiting in the queue.
These figures are typical rather than guaranteed, and vary with load and request size.