Which AI labs build models that best support user speech?
SpeechMap.AI tests how AI models respond to sensitive and controversial prompts. We measure what models refuse to say, redirect, or filter. Higher scores mean models engage more directly with difficult requests rather than declining or deflecting.
Labs are ranked by their Free Speech Index Score, a time-weighted average of models in each lab's latest release cycle (6 months, anchored to that lab's most recent release). Only labs with a release in the last 6 months are shown. For individual model results, see the Models page.
| Rank | Lab | Index | Peak Score | Models |
|---|---|---|---|---|
| #1 | Mistral AI | 91.0 | 98.2 | 7 |
| #2 | xAI | 82.2 | 98.2 | 8 |
| #3 | Google DeepMind | 77.4 | 88.0 | 6 |
| #4 | TNG Technology Consulting | 77.2 | 82.6 | 2 |
| #5 | Arcee AI | 74.7 | 82.2 | 2 |
| #6 | Zhipu AI | 73.2 | 85.8 | 9 |
| #7 | DeepSeek | 71.6 | 91.3 | 9 |
| #8 | Prime Intellect | 69.2 | 69.2 | 1 |
| #9 | xiaomi | 62.6 | 62.6 | 1 |
| #10 | inception | 56.5 | 56.5 | 1 |
| #11 | NVIDIA | 54.5 | 67.6 | 3 |
| #12 | Moonshot AI | 54.1 | 65.7 | 5 |
| #13 | Allen Institute for AI | 53.8 | 76.9 | 4 |
| #14 | Amazon | 51.2 | 65.8 | 2 |
| #15 | MiniMax | 51.0 | 55.2 | 3 |
| #16 | stepfun | 49.9 | 49.9 | 1 |
| #17 | liquid | 46.3 | 46.3 | 1 |
| #18 | OpenAI | 44.4 | 69.6 | 13 |
| #19 | Alibaba | 41.7 | 56.9 | 10 |
| #20 | Anthropic | 39.3 | 60.1 | 10 |
| #21 | ByteDance | 32.0 | 34.4 | 3 |