Our work
Research, evals and incident response
Research, evals and incident response all run on compute.
Research
Interpretability, targeted red-teaming, replication and new evaluation methods, run by academic and independent researchers.
Request research compute ↗Evals
Dangerous-capability evaluations before release, followed by reproduction, diagnosis and retesting.
Launch evaluations in Singapore ↗Incident response
Incident response is an emerging GPU need. When commercial AI services refused Hugging Face’s forensic analysis in July 2026, it ran an open-weight model on its own infrastructure.
Hugging Face incident report ↗Example: the open model frontier
Open models are critical for research, safety and affordable AI for all, but they come with risks: they can be downloaded, modified and run anywhere. Open model releases should be a moment of serious evaluation, and we exist to empower labs to launch more safely without compromising speed or capability.
Open-model releases, by month
View the numbers
| Month | 250+ | 1,000+ |
|---|---|---|
| 2024-09 | 12 | 5 |
| 2024-10 | 5 | 2 |
| 2024-11 | 3 | 2 |
| 2024-12 | 5 | 2 |
| 2025-01 | 10 | 4 |
| 2025-02 | 2 | 1 |
| 2025-03 | 4 | 1 |
| 2025-04 | 11 | 3 |
| 2025-05 | 4 | 4 |
| 2025-06 | 6 | 0 |
| 2025-07 | 7 | 4 |
| 2025-08 | 10 | 3 |
| 2025-09 | 14 | 4 |
| 2025-10 | 7 | 4 |
| 2025-11 | 6 | 0 |
| 2025-12 | 8 | 2 |
| 2026-01 | 11 | 5 |
| 2026-02 | 9 | 5 |
| 2026-03 | 5 | 3 |
| 2026-04 | 16 | 7 |
| 2026-05 | 5 | 2 |
| 2026-06 | 15 | 6 |
| 2026-07 | 7 | 0 |
| 2026-08 | 10 | 4 |
Put compute to work.
Bring a model, an evaluation, or the funding to make it happen.