Skip to main content
BenchLM
Data

Fastest LLMs by Output and First Answer

Celeris-1 (Celeris) is the fastest measured model at 1555 tokens/sec. Among models scoring 70 or higher, Claude Sonnet 5.5 leads at 138 tok/s. LFM2-24B-A2B has the lowest first-answer latency at 0.42s.

Runtime feed verified · Artificial Analysis

The fastest output model is not always the fastest model for a short request. Tokens/sec measures sustained generation after output begins. First-answer latency measures the opening wait. Sort both, keep a quality floor, and replay the request shape your application will use.

Median tokens/s, Latency first answer chunk (s). Measurements are screening evidence, not provider service-level guarantees.

Fastest Output

Celeris-1

1555 tok/s · Celeris

Lowest Latency

LFM2-24B-A2B

0.42s to first answer · LiquidAI

Fastest (Score 70+)

Claude Sonnet 5.5

138 tok/s · Score: 83.05

Top 15 — output speed (tok/s)

Ultra FastFastMediumSlow
  1. Celeris-11555 tok/s, Celeris, 0.54 seconds first-answer latency, score 27.14
  2. Mercury 2513 tok/s, Inception, 6.39 seconds first-answer latency, score 32.64
  3. Gemini 3.5 Flash-Lite461 tok/s, Google, 6.86 seconds first-answer latency, score 46.95
  4. Gemini 3.1 Flash-Lite376 tok/s, Google, 4.25 seconds first-answer latency, score 46.62
  5. Trinity-Large-Thinking367 tok/s, Arcee AI, 6.63 seconds first-answer latency, score 43.22
  6. Mercury 2.5366 tok/s, Inception, 5.69 seconds first-answer latency, score 32.89
  7. Mercury 2.5 Preview366 tok/s, Inception, 5.69 seconds first-answer latency
  8. Beam 1.0 501B-A23B345 tok/s, Reflection AI, 6.36 seconds first-answer latency, score 43.23
  9. LFM2.5-8B-A1B335 tok/s, LiquidAI, 7.62 seconds first-answer latency, score 23.86
  10. GPT-OSS 20B313 tok/s, OpenAI, 0.65 seconds first-answer latency, score 36.86
  11. Ling 3.0 Flash285 tok/s, InclusionAI, 9.81 seconds first-answer latency, score 45.1
  12. Gemini 3.7 Flash285 tok/s, Google, 8.52 seconds first-answer latency, score 65.91
  13. Nemotron 3 Nano 30B270 tok/s, NVIDIA, 8.33 seconds first-answer latency
  14. Claude Haiku 5.5241 tok/s, Anthropic, 343.38 seconds first-answer latency, score 67.64
  15. Nemotron 3 Nano Omni 30B A3B228 tok/s, NVIDIA, 9.53 seconds first-answer latency, score 31.4

Bar length is relative to the fastest model in the current filtered view. Exact median speed is shown at right; open a model for its complete evidence profile.

Average speed by provider

Inception

415 tok/s avg · 3 models

5.9s avg latency

LiquidAI

208 tok/s avg · 3 models

6.7s avg latency

InclusionAI

204 tok/s avg · 3 models

7.2s avg latency

NVIDIA

197 tok/s avg · 4 models

12s avg latency

Google

180 tok/s avg · 15 models

23s avg latency

StepFun

171 tok/s avg · 3 models

14.1s avg latency

Meta

128 tok/s avg · 8 models

21.7s avg latency

OpenAI

121 tok/s avg · 29 models

45.7s avg latency

Cohere

111 tok/s avg · 2 models

21.3s avg latency

xAI

99 tok/s avg · 9 models

19.3s avg latency

Mistral

93 tok/s avg · 9 models

4.4s avg latency

DeepSeek

89 tok/s avg · 8 models

11.4s avg latency

MiniMax

85 tok/s avg · 5 models

28.2s avg latency

Alibaba

82 tok/s avg · 16 models

41.8s avg latency

Anthropic

78 tok/s avg · 16 models

123.1s avg latency

Xiaomi

77 tok/s avg · 3 models

32.9s avg latency

Z.AI

64 tok/s avg · 9 models

30.7s avg latency

Moonshot AI

50 tok/s avg · 6 models

33.5s avg latency

Microsoft

33 tok/s avg · 3 models

2s avg latency

165 models match these filters.

RankModelSpeed / latency
1
Celeris-1Celeris · score 27.14
1555 tok/s0.54s latency
2
Mercury 2Inception · score 32.64
513 tok/s6.39s latency
6
Mercury 2.5Inception · score 32.89
366 tok/s5.69s latency
10
GPT-OSS 20BOpenAI · score 36.86
313 tok/s0.65s latency
17
InklingThinking Machines Lab · score 48.23
218 tok/s10.91s latency
24
Qwen3.7 MaxAlibaba · score 60.79
197 tok/s14.47s latency
25
Muse WonderMeta · score 67.38
197 tok/s35.22s latency
26
LFM2.5-2.6BLiquidAI · score 23.09
197 tok/s12.02s latency
29
Sarvam 105BSarvam · score 29.37
191 tok/s12.11s latency
35
Quasar 438BMultiverse Computing · score 49.05
180 tok/s12.39s latency
39
o3OpenAI · score 47.97
164 tok/s6.84s latency
40
GPT-4oOpenAI · score 25.74
160 tok/s0.91s latency
41
o3-miniOpenAI · score 38.5
160 tok/s7.12s latency
42
GPT-4.1OpenAI · score 36.26
155 tok/s0.89s latency
45
Command A+Cohere · score 28.62
152 tok/s13.64s latency
57
GPT-5 nanoOpenAI · score 36.89
137 tok/s83.3s latency
61
Nova ProAmazon · score 21.32
129 tok/s1.06s latency
63
Solar Pro 4Upstage · score not scored
121 tok/s18.19s latency
67
GPT-5.1OpenAI · score 56.39
117 tok/s28.57s latency
70
Grok 4.20xAI · score 55.27
108 tok/s17.04s latency
71
MiniMax M3MiniMax · score 49.41
107 tok/s19.73s latency
72
Grok 4.3xAI · score 52.58
105 tok/s24.18s latency
74
Grok 4.7xAI · score not scored
101 tok/s40.11s latency
76
GPT-5.5OpenAI · score 66.01
98 tok/s37.05s latency
78
o1OpenAI · score 35.8
98 tok/s32.29s latency
81
GPT-5.6 SolOpenAI · score 72.48
94 tok/s100.66s latency
82
GPT-5 (high)OpenAI · score not scored
94 tok/s1.04s latency
83
LFM2-24B-A2BLiquidAI · score not scored
92 tok/s0.42s latency
85
GPT-5.4OpenAI · score 65.43
89 tok/s157.51s latency
89
GLM-5.3Z.AI · score 60.96
86 tok/s25.62s latency
90
GPT-5 miniOpenAI · score 47.98
86 tok/s65.32s latency
91
Hy3Tencent · score not scored
84 tok/s26.41s latency
94
GLM-5.2Z.AI · score 59.71
82 tok/s27.9s latency
96
GPT-5.2OpenAI · score 58.86
80 tok/s81.31s latency
99
Qwen3.5-27BAlibaba · score 47.93
78 tok/s31.49s latency
100
Grok 4xAI · score 52.02
78 tok/s11.6s latency
101
LongCat-2.0Meituan · score not scored
76 tok/s29.14s latency
103
GPT-5.4 ProOpenAI · score 66.96
74 tok/s151.79s latency
104
GPT-6 AstraOpenAI · score 85.65
73 tok/s250.99s latency
105
GLM-5.1Z.AI · score 55.47
73 tok/s53.77s latency
106
GLM-5Z.AI · score 54.21
71 tok/s45.24s latency
117
Qwen3 MaxAlibaba · score 42.04
64 tok/s2.21s latency
120
Grok 4.6xAI · score 66.24
61 tok/s35.43s latency
122
Grok 4.5xAI · score 62.7
60 tok/s9.9s latency
126
Kimi K2.5Moonshot AI · score 52.55
58 tok/s52.9s latency
128
Qwen3.6 PlusAlibaba · score 53.41
56 tok/s100.68s latency
129
Qwen3.6-27BAlibaba · score 48.03
56 tok/s105.68s latency
132
Kimi K2Moonshot AI · score 33.37
54 tok/s1.49s latency
134
Qwen3 32BAlibaba · score not scored
53 tok/s1.23s latency
136
GLM-4.5Z.AI · score not scored
51 tok/s1.45s latency
139
Kimi K2.6Moonshot AI · score 56.64
44 tok/s2.81s latency
142
Qwen3.8 MaxAlibaba · score 63.57
41 tok/s52.16s latency
143
Phi-4Microsoft · score 22.87
41 tok/s2.64s latency
145
Qwen3.8-27BAlibaba · score 54.04
40 tok/s54.01s latency
147
Kimi K3Moonshot AI · score 68.42
38 tok/s56.25s latency
148
GLM-4.7Z.AI · score 47.74
38 tok/s55.52s latency
150
DeepSeek V3DeepSeek · score 31.53
37 tok/s1.25s latency
151
Gemma 4 31BGoogle · score 28.81
35 tok/s50.03s latency
152
DeepSeek-R1DeepSeek · score 40.66
35 tok/s57.8s latency
153
MiMo-V2.5Xiaomi · score not scored
35 tok/s63.46s latency
155
GLM-4.6Z.AI · score 41.92
34 tok/s4.57s latency
156
GPT-4o miniOpenAI · score 24.84
33 tok/s3.16s latency
157
GPT-4 TurboOpenAI · score 21.19
32 tok/s3.01s latency
158
Gemma 3 27BGoogle · score 28.41
31 tok/s2.04s latency
163
o3-proOpenAI · score 44.79
27 tok/s84.93s latency
164
Qwen3.5-4BAlibaba · score not scored
20 tok/s102.46s latency

Runtime rows use the same cross-provider feed. Prompt shape, region, provider load, account tier, cache state, and reasoning settings can change production results.

Two clocks choose different winners

First-answer latency dominates short interactive work: routing, classification, a tool argument, or the opening phrase of a voice turn. Sustained output dominates long reports and code generation after the response has started.

A rough completion estimate is first-answer time plus requested output tokens divided by tokens per second. It will not capture jitter, stop time, reasoning tokens, or retries, but it reveals when a slower starter can overtake a slower writer.

Read the TTFT explainer for timing boundaries, prompt prefill, and a reproducible measurement plan.

Keep speed behind a quality gate

A model that returns the wrong tool arguments quickly is not a fast production system. Build a task suite first, then identify the lowest-latency candidates that clear it. The score column is a broad screening signal, not a substitute for that application test.

Compare at least p50 and p95 latency from the deployment region. Hold prompt, output length, concurrency, API route, and reasoning budget constant. Keep failures and rate limits in the report so the surviving requests do not create a flattering distribution.

For voice, measure end of user speech to first audible output across transcription, model, speech generation, network, and playback. Text first-answer latency is only one segment.

Questions

What does tokens per second mean for LLMs?

Tokens per second measures sustained output after generation begins. A higher rate shortens long answers, but it does not reduce the initial wait by itself. Tokenization also differs across model families, so use the rate to screen candidates and confirm total completion time on the prompts your application sends.

What does the latency column measure?

Latency is the cross-provider runtime feed’s time from request to the first usable answer chunk. Lower is better within that test. For reasoning models it can include work performed before visible output. It is not full voice latency, browser render time, or a service-level guarantee for another region and account.

Which LLM is the fastest?

Currently, Celeris-1 by Celeris is the fastest at 1555 tokens/second. The fastest model scoring above 70 overall is Claude Sonnet 5.5 at 138 tok/s.

Why are reasoning models slower?

Some reasoning settings perform additional work before a usable answer appears, which can increase first-answer latency and billed output. The effect varies by model, provider, prompt, and reasoning budget. Measure the exact setting you will deploy, then compare task success with latency and cost rather than assuming every reasoning label behaves alike.

Know when it’s worth switching models

The model to choose, the cheaper alternative, and the release we would wait on.

Read a sample issue

Join 7,000+ subscribers.

One email each week. Unsubscribe anytime.