Speed · tracked daily
The fastest AI models right now
The fastest AI model right now is Qwen: Qwen3.8 27B (batch) at 835 tokens/second via Cerebras. Speed is tracked daily, not benchmarked once - providers re-provision constantly, so a static "fastest model" table goes stale within days.
- #1 Qwen: Qwen3.8 27B (batch) - 835 tok/s via Cerebras
- #2 OpenAI: gpt-oss-120b (batch) - 813 tok/s via Cerebras
- #3 Z.ai: GLM 5.2 - 528 tok/s via Decart
- #4 OpenAI: gpt-oss-20b (batch) - 453 tok/s via Groq
- #5 NVIDIA: Nemotron 3.5 Super VL (free) - 310 tok/s via Venice
- #6 Google: Gemma 4 31B (free) - 304 tok/s via Cerebras
- #7 NVIDIA: Nemotron 3.5 Lightning (free) - 269 tok/s via CoreWeave
- #8 MiniMax: MiniMax M2.7 - 228 tok/s via Groq
- #9 DeepSeek: DeepSeek V4 Flash 0423 - 213 tok/s via Wafer
- #10 inclusionAI: Ling 3.0 Flash Fin - 213 tok/s via Novita
- #11 Z.ai: GLM 5.3 (batch) - 201 tok/s via ModelRun
- #12 MiniMax: MiniMax M3.1 Flash - 199 tok/s via DigitalOcean
- #13 MiniMax: MiniMax M3 (batch) - 196 tok/s via CoreWeave
- #14 Qwen: Qwen3.6 35B A3B - 196 tok/s via Phala
- #15 DeepSeek: DeepSeek V4.1 Flash (batch) - 193 tok/s via Together
| # | Model | Fastest provider | Throughput | Latency p50 | Quality | Quality per $ |
|---|---|---|---|---|---|---|
| 01 | Qwen: Qwen3.8 27B (batch) | Cerebras | 835 tok/s | 342ms | 20.1 | 34.6 pts/$ |
| 02 | OpenAI: gpt-oss-120b (batch) | Cerebras | 813 tok/s | 154ms | 11.6 | 692.5 pts/$ |
| 03 | Z.ai: GLM 5.2 | Decart | 528 tok/s | 417ms | - | - |
| 04 | OpenAI: gpt-oss-20b (batch) | Groq | 453 tok/s | 546ms | 9 | 163.6 pts/$ |
| 05 | NVIDIA: Nemotron 3.5 Super VL (free) | Venice | 310 tok/s | 500ms | 13.6 | 57.3 pts/$ |
| 06 | Google: Gemma 4 31B (free) | Cerebras | 304 tok/s | 244ms | 4.4 | 17.6 pts/$ |
| 07 | NVIDIA: Nemotron 3.5 Lightning (free) | CoreWeave | 269 tok/s | 197ms | 5.5 | 53.7 pts/$ |
| 08 | MiniMax: MiniMax M2.7 | Groq | 228 tok/s | 267ms | - | - |
| 09 | DeepSeek: DeepSeek V4 Flash 0423 | Wafer | 213 tok/s | 265ms | - | - |
| 10 | inclusionAI: Ling 3.0 Flash Fin | Novita | 213 tok/s | 1296ms | - | - |
| 11 | Z.ai: GLM 5.3 (batch) | ModelRun | 201 tok/s | 140ms | 31.7 | 78.8 pts/$ |
| 12 | MiniMax: MiniMax M3.1 Flash | DigitalOcean | 199 tok/s | 1110ms | 29.1 | 110.9 pts/$ |
| 13 | MiniMax: MiniMax M3 (batch) | CoreWeave | 196 tok/s | 331ms | 18.9 | 36.0 pts/$ |
| 14 | Qwen: Qwen3.6 35B A3B | Phala | 196 tok/s | 1778ms | - | - |
| 15 | DeepSeek: DeepSeek V4.1 Flash (batch) | Together | 193 tok/s | 208ms | 28.6 | 109.0 pts/$ |
| 16 | Anthropic: Claude Haiku 5.5 (batch) | Claude Platform on AWS | 189 tok/s | 2463ms | 36.5 | 182.5 pts/$ |
| 17 | Google: Gemini 2.5 Flash Lite | 181 tok/s | 511ms | - | - | |
| 18 | Google: Gemini 3.1 Flash Lite (batch) | Google AI Studio | 178 tok/s | 597ms | - | - |
| 19 | StepFun: Step 5 Preview | InferenceNet | 168 tok/s | 762ms | 28.5 | 20.0 pts/$ |
| 20 | Thinking Machines: Inkling | BaseTen | 164 tok/s | 333ms | 12.5 | 7.1 pts/$ |
| 21 | Z.ai: GLM 5.3 Flash (batch) | CoreWeave | 151 tok/s | 544ms | 27.8 | 117.1 pts/$ |
| 22 | Google: Gemini 2.5 Flash | Google AI Studio | 145 tok/s | 497ms | - | - |
| 23 | Google: Gemma 4 26B A4B (free) | Makora | 143 tok/s | 278ms | - | - |
| 24 | OpenAI: GPT-6 Luna (batch) | Amazon Bedrock | 143 tok/s | 1050ms | 27.9 | 139.5 pts/$ |
| 25 | OpenAI: GPT-5.6 Sol (batch) | OpenAI | 131 tok/s | 5168ms | - | - |
| 26 | Google: Gemini 3.5 Flash (batch) | 131 tok/s | 1513ms | - | - | |
| 27 | Mistral: Mistral Nemo | Mistral | 130 tok/s | 209ms | - | - |
| 28 | Meta: Muse Spark 1.2 Contributor | Meta | 126 tok/s | 2473ms | - | - |
| 29 | DeepSeek: DeepSeek V4 Pro 0423 | BaseTen | 122 tok/s | 531ms | 23.1 | 18.4 pts/$ |
| 30 | MoonshotAI: Kimi K3 (batch) | Modal | 121 tok/s | 501ms | 17.4 | 4.2 pts/$ |
| 31 | OpenAI: GPT-5.6 Luna (batch) | Amazon Bedrock | 115 tok/s | 1262ms | 23.8 | 52.9 pts/$ |
| 32 | NVIDIA: Nemotron 3 Super | Crusoe | 115 tok/s | 367ms | 12.8 | 81.9 pts/$ |
| 33 | NVIDIA: Nemotron 3 Ultra (free) | BaseTen | 114 tok/s | 273ms | 11.8 | 11.2 pts/$ |
| 34 | Poolside: Laguna XS 2.1 (free) | Poolside | 114 tok/s | 268ms | - | - |
| 35 | Qwen: Qwen3.8 Flash | Makora | 112 tok/s | 1011ms | - | - |
| 36 | Anthropic: Claude Sonnet 5.5 (batch) | 111 tok/s | 1564ms | 54.5 | 13.6 pts/$ | |
| 37 | OpenAI: GPT-5.6 Terra (batch) | Amazon Bedrock | 109 tok/s | 699ms | 32.3 | 7.2 pts/$ |
| 38 | Xiaomi: MiMo-V2.6-Pro | PrimaLabs | 109 tok/s | 1103ms | 29.9 | 55.0 pts/$ |
| 39 | DeepSeek: DeepSeek V3.2 | Mara | 108 tok/s | 1261ms | - | - |
| 40 | Meta: Muse Spark 1.2 | Meta | 105 tok/s | 2525ms | - | - |
Top 5, speed over time
- Qwen: Qwen3.8 27B (batch)
- OpenAI: gpt-oss-120b (batch)
- Z.ai: GLM 5.2
- OpenAI: gpt-oss-20b (batch)
- NVIDIA: Nemotron 3.5 Super VL (free)
Frequently asked questions
What is the fastest AI model?
As of the latest data, Qwen: Qwen3.8 27B (batch) is the fastest widely available AI model at 835 tokens/second via Cerebras.
How fast is Qwen: Qwen3.8 27B (batch) in tokens per second?
Qwen: Qwen3.8 27B (batch) generates approximately 835 tokens per second on its fastest available provider (Cerebras), measured directly from OpenRouter's endpoint data.