Skip to main content
Radar

Five or fewer confirmed AI changes, with original sources, on mornings when something changed.A free source-linked morning brief.

Start free brief
BenchLM

Fastest LLMs by Output and First Answer

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.

Cross-provider runtime snapshot last updated 2026-08-14. Median tokens/s, Latency first answer chunk (s). Measurements are screening evidence, not provider service-level guarantees.

Fastest Output

Celeris-1

1560 tok/s · Celeris

Lowest Latency

LFM2-24B-A2B

0.42s to first answer · LiquidAI

Fastest (Score 70+)

Gemini 3.6 Flash

225 tok/s · Score: 75.54

Top 15 — Output Speed (tok/s)

Ultra FastFastMediumSlow
  1. Celeris-11560 tok/s, Celeris, 0.6 seconds first-answer latency
  2. Mercury 2820 tok/s, Inception, 3.34 seconds first-answer latency, score 48.14
  3. Step 3.7 Flash399 tok/s, StepFun, 5.85 seconds first-answer latency, score 50.91
  4. Ling 3.0 Flash374 tok/s, InclusionAI, 7.4 seconds first-answer latency, score 53.78
  5. Gemini 3.5 Flash-Lite368 tok/s, Google, 10.11 seconds first-answer latency, score 65.29
  6. Gemini 3.7 Flash340 tok/s, Google, 9.83 seconds first-answer latency, score 61.36
  7. LFM2.5-8B-A1B340 tok/s, LiquidAI, 7.99 seconds first-answer latency, score 41.62
  8. Gemini 3.1 Flash-Lite335 tok/s, Google, 5.54 seconds first-answer latency, score 50.51
  9. Nemotron 3 Nano Omni 30B A3B325 tok/s, NVIDIA, 7.16 seconds first-answer latency, score 44.4
  10. GPT-OSS 20B313 tok/s, OpenAI, 0.65 seconds first-answer latency, score 42.2
  11. Nemotron 3 Nano 30B286 tok/s, NVIDIA, 8.12 seconds first-answer latency, score 53.36
  12. Gemini 3.6 Flash225 tok/s, Google, 19.69 seconds first-answer latency, score 75.54
  13. Muse Spark 1.1217 tok/s, Meta, 12.12 seconds first-answer latency, score 76.88
  14. Gemini 2.5 Flash215 tok/s, Google, 0.5 seconds first-answer latency, score 47.45
  15. Qwen3.7 Max208 tok/s, Alibaba, 13.88 seconds first-answer latency, score 71.6

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

StepFun

298 tok/s avg · 2 models

8.6s avg latency

InclusionAI

250 tok/s avg · 2 models

4.3s avg latency

NVIDIA

224 tok/s avg · 4 models

11.8s avg latency

LiquidAI

216 tok/s avg · 2 models

4.2s avg latency

Google

189 tok/s avg · 12 models

17.3s avg latency

Meta

123 tok/s avg · 4 models

4s avg latency

OpenAI

115 tok/s avg · 27 models

48.9s avg latency

Xiaomi

115 tok/s avg · 2 models

11.9s avg latency

xAI

99 tok/s avg · 8 models

17.4s avg latency

Alibaba

98 tok/s avg · 11 models

41s avg latency

Mistral

98 tok/s avg · 7 models

2.9s avg latency

MiniMax

82 tok/s avg · 3 models

29s avg latency

Z.AI

76 tok/s avg · 8 models

29.2s avg latency

DeepSeek

72 tok/s avg · 3 models

7.5s avg latency

Anthropic

54 tok/s avg · 12 models

33.8s avg latency

Moonshot AI

40 tok/s avg · 5 models

37.6s avg latency

Microsoft

27 tok/s avg · 2 models

1.5s avg latency

122 models match these filters.

RankModelSpeed / latency
1
Celeris-1Celeris · score not scored
1560 tok/s0.6s latency
2
Mercury 2Inception · score 48.14
820 tok/s3.34s latency
10
GPT-OSS 20BOpenAI · score 42.2
313 tok/s0.65s latency
15
Qwen3.7 MaxAlibaba · score 71.6
208 tok/s13.88s latency
18
Command A+Cohere · score 47.56
199 tok/s10.46s latency
27
o3-miniOpenAI · score 46.68
160 tok/s7.12s latency
34
GPT-4oOpenAI · score 40.81
141 tok/s0.83s latency
37
GPT-5 nanoOpenAI · score 46.49
137 tok/s83.3s latency
39
Grok 4.3xAI · score 63.95
133 tok/s24.37s latency
45
GPT-5.4OpenAI · score 73.36
121 tok/s144.51s latency
48
o3OpenAI · score 47.01
118 tok/s7.37s latency
51
GLM-4.7Z.AI · score 60.66
110 tok/s19.25s latency
53
GLM-5.2Z.AI · score 63.34
108 tok/s19.89s latency
55
MiMo-V2.5Xiaomi · score 59.16
101 tok/s21.74s latency
56
GPT-4.1OpenAI · score 50.84
101 tok/s1.02s latency
57
o1OpenAI · score 48.17
98 tok/s32.29s latency
60
Nova ProAmazon · score 19.56
96 tok/s1.11s latency
61
Grok 4.20xAI · score 55.19
95 tok/s27.81s latency
64
GPT-5.1OpenAI · score 53.54
88 tok/s41.06s latency
65
MiniMax M3MiniMax · score 68.74
86 tok/s25s latency
66
GPT-5 miniOpenAI · score 43.08
86 tok/s65.32s latency
70
Qwen3.5-27BAlibaba · score 59.82
77 tok/s31.44s latency
71
GPT-5.4 ProOpenAI · score 61.26
74 tok/s151.79s latency
73
Grok 3 MinixAI · score not scored
74 tok/s27.79s latency
74
InklingThinking Machines Lab · score 67.04
73 tok/s29.1s latency
75
GPT-5.2OpenAI · score 58.24
73 tok/s160.9s latency
76
GLM-5.1Z.AI · score 66.91
69 tok/s56.24s latency
78
GPT-5.5OpenAI · score 73.37
67 tok/s101.36s latency
80
Grok 4.6xAI · score 63.41
66 tok/s32.3s latency
82
GLM-4.6Z.AI · score 54.34
66 tok/s2.68s latency
83
Hy3Tencent · score 68.02
64 tok/s34.24s latency
86
Qwen3 MaxAlibaba · score 48.26
62 tok/s2.36s latency
87
Grok 4.5xAI · score 75.42
60 tok/s8.14s latency
92
Qwen3.6-27BAlibaba · score 53.73
55 tok/s107.72s latency
93
Grok 4xAI · score 59.61
54 tok/s15.6s latency
96
GLM-4.5Z.AI · score 58.02
51 tok/s1.45s latency
97
Kimi K2.5Moonshot AI · score 58.89
48 tok/s65.23s latency
98
Qwen3.8 MaxAlibaba · score 79.91
47 tok/s45.03s latency
103
GLM-5Z.AI · score 65.61
40 tok/s79.77s latency
105
LongCat-2.0Meituan · score not scored
40 tok/s52.71s latency
106
Kimi K3Moonshot AI · score 80.5
39 tok/s54.29s latency
108
Kimi K2.6Moonshot AI · score 60.21
39 tok/s2.9s latency
109
Kimi K2Moonshot AI · score 26.32
38 tok/s1.44s latency
112
Gemma 4 31BGoogle · score 60.15
36 tok/s49.9s latency
114
Phi-4Microsoft · score 22.41
35 tok/s2.24s latency
115
GPT-4o miniOpenAI · score 37.22
33 tok/s3.16s latency
116
Gemma 3 27BGoogle · score 41.04
31 tok/s2.04s latency
117
GPT-4 TurboOpenAI · score 26.71
31 tok/s3.56s latency
121
o3-proOpenAI · score 47.43
27 tok/s84.93s 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.

Price vs Performance

See which models offer the best value

Full Pricing Table

Compare all LLM API prices

Benchmark Confidence

Which scores can you trust?

Frequently Asked 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 1560 tokens/second. The fastest model scoring above 70 overall is Gemini 3.6 Flash at 225 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 2,000+ readers.

One email each week. Unsubscribe anytime.