Latency: time to first answer
Median seconds until the answer starts, including any reasoning phase (Artificial Analysis), or time to first streamed token where that is all we have.
As of 24 Oct 2026, BenchLeader ranks Llama 3.2 1B Instruct first on the Latency: time to first answer board at 0.23 s, ahead of Qwen2.5 7B Instruct (0.27 s) and gpt-oss-20b (0.29 s).
665 models · data as of 24 Oct 2026
- 1Llama 3.2 1B Instruct0.23 s
- 2Qwen2.5 7B Instruct0.27 s
- 3gpt-oss-20b0.29 s
- 4Gemini 2.5 Flash-Lite (none)0.33 s
- 5Llama 3.2 3B Instruct0.35 s
- 6Beam 1.0 501B A23B0.36 s
- 7Gemini 2.5 Flash (none)0.36 s
- 8Nemotron 3 Ultra 550B A55B0.36 s
- 9Qwen3 32B0.38 s
- 10GLM 4.7 Flash0.39 s
- 11Palmyra X50.41 s
- 12Hermes 4 405B (none)0.43 s
- 13gpt-oss-120b0.44 s
- 14Nemotron 3 Nano 30B A3B0.45 s
- 15Qwen-Plus0.46 s
264 of 264
| # | |||
|---|---|---|---|
| 1 |  DeepSeek V4 Flash 0420maxopen | 119.66 s | 47 tok/s |
| 2 |  Qwen3.6 Plus | 101.62 s | 56 tok/s |
| 3 | DeepSeek R1 0528open ↗ | 62.10 s | 33 tok/s |
| 4 |  Seed-OSS-36B-Instructopen ↗ | 55.55 s | 38 tok/s |
| 5 | Qwen3.8 2.4T A95Bopen ↗ | 49.49 s | 43 tok/s |
| 6 |  MiniMax M2.7open ↗ | 46.82 s | 55 tok/s |
| 7 |  Kimi K2.7 Codeopen ↗ | 44.41 s | 54 tok/s |
| 8 | Qwen3.8-Flash-Nextopen ↗ | 40.85 s | 52 tok/s |
| 9 |  MiMo-V2.6-Proopen ↗ | 39.78 s | 54 tok/s |
| 10 | MiMo-V2.6-Flashopen ↗ | 37.78 s | 60 tok/s |
| 11 |  Ling 3.0 Tinyopen | 37.78 s | 57 tok/s |
| 12 | Qwen3.7 Plus | 36.06 s | 59 tok/s |
| 13 |  North Mini Codeopen | 34.37 s | 59 tok/s |
| 14 |  GLM 5.3 Flashopen ↗ | 32.42 s | 69 tok/s |
| 15 |  Grok Build 0.1 | 32.25 s | 65 tok/s |
| 16 | Grok 3 minithinking | 31.95 s | 66 tok/s |
| 17 | MiniMax-M2.1open ↗ | 31.11 s | 68 tok/s |
| 18 | MiniMax M2.5open ↗ | 30.95 s | 68 tok/s |
| 19 | GLM-4.5-Airopen ↗ | 29.46 s | 74 tok/s |
| 20 |  LongCat 2.0open ↗ | 28.74 s | 77 tok/s |
| 21 |  Sonar Reasoning Prothinking | 28.00 s | 20 tok/s |
| 22 | NewKimi K3.1 Pro | 27.74 s | 81 tok/s |
| 23 | MiMo-V2.5open ↗ | 27.35 s | 76 tok/s |
| 24 |  Magistral Small 1.2open | 27.02 s | 78 tok/s |
| 25 |  o4-minihigh | 25.08 s | 149 tok/s |
| 26 | MiniMax M3open ↗ | 25.02 s | 84 tok/s |
| 27 | Mistral Large 4open | 24.75 s | 93 tok/s |
| 28 |  Reka Flash 3open ↗ | 24.17 s | 88 tok/s |
| 29 |  Gemini 2.5 Pro | 23.69 s | 121 tok/s |
| 30 | MiniMax-M1open | 23.03 s | 94 tok/s |
| 31 | NewMiniMax M3.1 Flashopen ↗ | 22.41 s | 96 tok/s |
| 32 |  Solar Mini 4 | 22.32 s | 98 tok/s |
| 33 | MiniMax-M2open ↗ | 22.18 s | 96 tok/s |
| 34 |  Granite 4.2 8Bopen ↗ | 21.24 s | 97 tok/s |
| 35 | Gemini 3.1 Pro | 21.06 s | 132 tok/s |
| 36 | Granite 4.2 30Bopen ↗ | 20.88 s | 99 tok/s |
| 37 | Ring-2.6-1Topen | 20.74 s | 118 tok/s |
| 38 | DeepSeek V3.1thinkingopen ↗ | 19.33 s | 120 tok/s |
| 39 |  Agnes 2.5 Pro Beta | 19.09 s | 113 tok/s |
| 40 | Granite 4.2 3Bopen ↗ | 19.02 s | 110 tok/s |
| 41 | Kimi K2 Thinkingthinkingopen ↗ | 18.43 s | 119 tok/s |
| 42 | Solar Pro 4 | 18.36 s | 121 tok/s |
| 43 | DeepSeek R1open ↗ | 17.39 s | 144 tok/s |
| 44 |  Step 3.5 Flashopen ↗ | 16.90 s | 146 tok/s |
| 45 | Step 3.5 Flash 0202open | 16.79 s | 148 tok/s |
| 46 | Solar Pro 3 | 15.83 s | 137 tok/s |
| 47 | Mistral Medium 3.5open | 14.83 s | 159 tok/s |
| 48 | Qwen3.7max | 14.12 s | 201 tok/s |
| 49 | Command A+open | 13.85 s | 150 tok/s |
| 50 | GPT-5.5 Instant | 13.73 s | 156 tok/s |
| 51 | Step 3.7 Flashopen ↗ | 13.61 s | 179 tok/s |
| 52 | Granite 3.3 8Bno reasoningopen ↗ | 12.44 s | 38 tok/s |
| 53 |  Quasar 438Bmax | 12.00 s | 185 tok/s |
| 54 |  Sarvam 105Bhighopen | 11.76 s | 197 tok/s |
| 55 | GPT-5.5 Proxhigh | 11.41 s | 27 tok/s |
| 56 | Ling 3.1 Flashopen | 11.04 s | 220 tok/s |
| 57 | Inkling Smallopen ↗ | 10.54 s | 224 tok/s |
| 58 | Step 5 Previewopen | 10.13 s | 224 tok/s |
| 59 | GPT-5-Codexhigh | 10.09 s | 166 tok/s |
| 60 | GPT-6 Astra Promax | 9.88 s | 34 tok/s |
Cite as: BenchLeader, “Latency: time to first answer”, data as of 24 Oct 2026, https://www.benchleader.com/leaderboards/latency.