Best Models
Mercury 2 is currently the highest-ranked model for Speed in the available data.
Use the top-ranked model as a starting point, not an automatic decision.
Compare the top choices on price, speed, availability, and your own constraints.
Open the linked result page and methodology before quoting the ranking.
Median generated output tokens per second. Higher means the model streams completions faster after it starts responding.
| Rank | Model | Provider | Value |
|---|---|---|---|
| #1 | Mercury 2 | Inception | 880.1 tok/s |
| #2 | Granite 4.0 H Small | IBM | 439.5 tok/s |
| #3 | Granite 3.3 8B (Non-reasoning) | IBM | 414.9 tok/s |
| #4 | LFM2.5-VL-1.6B | Liquid AI | 392.7 tok/s |
| #5 | Step 3.7 Flash | StepFun | 385.2 tok/s |
This answer uses the exact same metric and live database ranking as the linked Easy Benchmarks leaderboard.
Source: Artificial Analysis Output Speed. Updated: Jul 16, 2026.
| #6 | HyperNova 60B 2605 | Multiverse Computing | 350.9 tok/s |
| #7 | LFM2.5-8B-A1B | Liquid AI | 343.4 tok/s |
| #8 | Nemotron 3 Nano Omni 30B A3B Reasoning | NVIDIA | 320 tok/s |
| #9 | gpt-oss-120b (low) | OpenAI | 296.3 tok/s |
| #10 | Gemini 3.1 Flash-Lite | 291.6 tok/s |
| #11 | Llama 3.1 Nemotron Instruct 70B | NVIDIA | 286.3 tok/s |
| #12 | NVIDIA Nemotron Nano 12B v2 VL (Reasoning) | NVIDIA | 282.5 tok/s |
| #13 | gpt-oss-20b (low) | OpenAI | 260.9 tok/s |
| #14 | Step 3.5 Flash 2603 | StepFun | 256.9 tok/s |
| #15 | Nova Micro | Amazon | 249.8 tok/s |
| #16 | Step 3.5 Flash | StepFun | 247.4 tok/s |
| #17 | GPT-5.6 Luna (max) | OpenAI | 243.6 tok/s |
| #18 | Gemini 3.5 Flash (high) | 241.2 tok/s |
| #19 | Gemini 2.5 Flash-Lite (Reasoning) | 237.8 tok/s |
| #20 | Gemini 3.5 Flash (medium) | 234.7 tok/s |