Speed: output tokens per second

Median output throughput measured across healthy providers over the last 30 minutes of live traffic (OpenRouter).

As of 21 Oct 2026, BenchLeader ranks Celeris-1 first on the Speed: output tokens per second board at 1533, ahead of K2 Horizon 375B A23B (1115) and Mercury 2.5 (781).

661 models · data as of 21 Oct 2026

1. 1Celeris-11533
2. 2K2 Horizon 375B A23B1115
3. 3Mercury 2.5781
4. 4Mercury 2499
5. 5Gemini 3.5 Flash Lite454
6. 6DeepSeek V4.1 Flash (none)442
7. 7Nemotron 3.5 Super VL428
8. 8Gemini 2.5 Flash-Lite (thinking)422
9. 9Trinity Large (thinking)361
10. 10Grok 4.20 Multi-Agent356
11. 11Gemini 3.1 Flash Lite352
12. 12Beam 1.0345
13. 13Muse Spark 1.2 (xhigh)315
14. 14Ling 3.0 Flash312
15. 15Nemotron 3.5 Lightning312

SearchAll providersAlibabaAmazonAnt GroupAnthropicArcee AIByteDanceCelerisCohereDeepSeekGoogleIBMInceptionMBZUAIMeituanMetaMicrosoftMiniMaxMistral AIMoonshot AIMultiverse ComputingNous ResearchNVIDIAOpenAIPerplexityReflectionRekaSapiens AISarvamSpaceXAIStepFunSwiss AITencentThinking MachinesUnknownUpstageWriterXiaomiZhipu AIOpen weightsNew modelsBest per model (264)264 of 264

#	Model	tok/s ↓	First token ↕
1	CCeleris-1Celeris	1533	0.6 s
2	K2 Horizon 375B A23BMBZUAIopen	1115	–
3	Mercury 2.5Inception	781	1.6 s
4	Mercury 2Inception	499	1.1 s
5	Gemini 3.5 Flash LiteGoogle	454	0.6 s
6	DeepSeek V4.1 Flashno reasoningDeepSeekopen ↗	442	0.8 s
7	NewNemotron 3.5 Super VLNVIDIAopen ↗	428	0.7 s
8	Gemini 2.5 Flash-LitethinkingGoogle	422	–
9	Trinity LargethinkingArcee AIopen	361	0.9 s
10	Grok 4.20 Multi-AgentSpaceXAI	356	5.5 s
11	Gemini 3.1 Flash LiteGoogle	352	0.8 s
12	NewBeam 1.0Reflectionopen	345	–
13	Muse Spark 1.2xhighMeta	315	2.8 s
14	Ling 3.0 FlashAnt Groupopen ↗	312	0.9 s
15	Nemotron 3.5 LightningNVIDIAopen ↗	312	0.2 s
16	Ling 3.0 Flash FinAnt Groupopen	283	–
17	Muse Spark 1.3xhighMeta	282	3.0 s
18	Muse Spark 1.1xhighMeta	280	1.7 s
19	Nova MicroAmazon	280	0.4 s
20	GPT-5.4 MinixhighOpenAI	276	1.0 s
21	gpt-oss-20blowOpenAIopen ↗	266	0.3 s
22	gpt-oss-120blowOpenAIopen ↗	264	0.4 s
23	Nemotron 3 Nano 30B A3Bno reasoningNVIDIAopen ↗	262	0.6 s
24	GPT-4.1 nanoOpenAI	249	0.8 s
25	Qwen3.5 Omni FlashAlibaba	239	1.9 s
26	Nemotron 3 Nano Omni 30B A3BthinkingNVIDIAopen	238	–
27	Claude Haiku 5.5maxAnthropic	234	1.0 s
28	o3-miniOpenAI	229	3.1 s
29	Gemini 2.5 FlashthinkingGoogle	228	–
30	Step 3.7 FlashStepFunopen ↗	223	3.6 s
31	Gemini 3 Flashno reasoningGoogle	217	0.9 s
32	Ling 3.1 FlashAnt Groupopen	216	1.2 s
33	Gemini 3.6 FlashhighGoogle	214	1.1 s
34	Sarvam 105BhighSarvamopen	209	–
35	NewDeepSeek V4.1 Flash 1017maxDeepSeekopen ↗	209	0.9 s
36	Gemini 3.5 FlashhighGoogle	208	1.6 s
37	Nemotron Nano 12B v2 VLno reasoningNVIDIAopen	203	1.1 s
38	Deepseek v4 Flash VisionmaxDeepSeek	202	1.3 s
39	Qwen3 7maxAlibaba	201	1.1 s
40	Mistral Small 4thinkingMistral AIopen	196	0.4 s
41	Nova 2.0 LitemediumAmazon	192	–
42	Ministral 3 3BMistral AIopen ↗	192	0.3 s
43	GPT-5-CodexhighOpenAI	190	–
44	TInkling SmallThinking Machinesopen ↗	190	0.4 s
45	Quasar 438BmaxMultiverse Computing	183	–
46	GPT-4oOpenAI	182	0.6 s
47	DeepSeek R1DeepSeekopen ↗	180	–
48	GLM 5.2no reasoningZhipu AIopen ↗	180	1.1 s
49	Llama 3.2 3B InstructMetaopen ↗	176	0.3 s
50	GPT-5.4 NanomediumOpenAI	172	0.9 s
51	GPT-4.1 minihighOpenAI	172	0.7 s
52	GPT-5 nanohighOpenAI	172	1.8 s
53	Nova LiteAmazon	170	0.6 s
54	GPT-5.1-Codex-MinihighOpenAI	167	4.7 s
55	NVIDIA Nemotron Nano 9B v2no reasoningNVIDIAopen	167	1.1 s
56	Qwen3.5-35B-A3Bno reasoningAlibabaopen ↗	167	2.1 s
57	Muse WondermaxMeta	161	3.0 s
58	Command A+Cohereopen	160	0.3 s
59	Kimi K2 ThinkingthinkingMoonshot AIopen ↗	159	1.2 s
60	o4-minihighOpenAI	158	3.7 s

Show 100 more

Cite as: BenchLeader, “Speed: output tokens per second”, https://www.benchleader.com/leaderboards/speed, data as of 21 Oct 2026.