Training compute growth
Maximum disclosed training compute per year on a log10(FLOP) scale. Missing values remain absent; the current year is incomplete.
Training compute growth
Total 471 · Page 2/19 · Sample synced 11/09, 03:30 UTC
| Name | Group | Reported value | Date | Notes | Source |
|---|---|---|---|---|---|
| Inflection-2 | Inflection AI | 10,010,000,000,000,000,000,000,000 FLOP | 2023-11-22 | Hosted access (no API) | Source ↗ |
| DeepSeek-V4-Pro | DeepSeek | 9,702,000,000,000,000,000,000,000 FLOP | 2026-04-24 | Open weights (unrestricted) | Source ↗ |
| Reka Core | Reka AI | 8,400,010,000,000,001,000,000,000 FLOP | 2024-04-15 | API access | Source ↗ |
| Inflection-2.5 | Inflection AI | 8,000,001,000,000,000,000,000,000 FLOP | 2024-03-07 | Hosted access (no API) | Source ↗ |
| Llama-3.1-Nemotron-70B-Instruct | Nvidia,Meta AI | 7,929,000,000,000,000,000,000,000 FLOP | 2024-10-12 | Open weights (restricted use) | Source ↗ |
| Llama 3-70B | Meta AI | 7,861,000,000,000,000,000,000,000 FLOP | 2024-04-18 | Open weights (restricted use) | Source ↗ |
| Qwen2.5 Instruct (72B) | Alibaba | 7,851,600,000,000,000,000,000,000 FLOP | 2024-09-19 | Open weights (restricted use) | Source ↗ |
| Qwen2.5-72B | Alibaba | 7,800,000,000,000,000,000,000,000 FLOP | 2024-09-19 | Open weights (unrestricted) | Source ↗ |
| PaLM 2 | 7,340,000,000,000,000,000,000,000 FLOP | 2023-05-10 | API access | Source ↗ | |
| Telechat2-115B | China Telecom | 6,900,000,000,000,000,000,000,000 FLOP | 2024-09-20 | Open weights (restricted use) | Source ↗ |
| Llama 3.3 70B | Meta AI | 6,864,976,800,000,000,000,000,000 FLOP | 2024-12-06 | Open weights (restricted use) | Source ↗ |
| GLM-5 | Z.ai (Zhipu AI) | 6,840,000,000,000,000,000,000,000 FLOP | 2026-02-12 | Open weights (unrestricted) | Source ↗ |
| MiMo-V2.5-Pro | Xiaomi Corp | 6,804,000,000,000,000,000,000,000 FLOP | 2026-04-23 | Open weights (unrestricted) | Source ↗ |
| Nemotron 3 Ultra | Nvidia | 6,600,000,000,000,000,000,000,000 FLOP | 2026-06-04 | Open weights (unrestricted) | Source ↗ |
| Amazon Nova Pro | Amazon | 6,000,010,000,000,001,000,000,000 FLOP | 2024-12-03 | API access | Source ↗ |
| Ling-1T | Ant Group | 6,000,001,000,000,000,000,000,000 FLOP | 2025-10-10 | Open weights (unrestricted) | Source ↗ |
| Kimi K2.5 | Moonshot | 5,800,000,000,000,000,000,000,000 FLOP | 2026-02-02 | Open weights (unrestricted) | Source ↗ |
| Hunyuan-TurboS | Tencent | 5,400,000,000,000,000,000,000,000 FLOP | 2025-03-11 | API access | Source ↗ |
| Qwen3-VL-235B-A22B | Alibaba | 5,038,044,000,000,000,000,000,000 FLOP | 2025-09-23 | Open weights (unrestricted) | Source ↗ |
| gpt-oss-120b | OpenAI | 4,940,000,000,000,000,000,000,000 FLOP | 2025-08-05 | Open weights (unrestricted) | Source ↗ |
| Amazon Titan | Amazon | 4,800,000,000,000,000,000,000,000 FLOP | 2023-09-28 | API access | Source ↗ |
| Qwen3-235B-A22B-Thinking (Jul 2025) | Alibaba | 4,752,000,000,000,000,000,000,000 FLOP | 2025-07-25 | Open weights (unrestricted) | Source ↗ |
| Qwen3-235B-A22B (Jul 2025) | Alibaba | 4,752,000,000,000,000,000,000,000 FLOP | 2025-07-25 | Open weights (unrestricted) | Source ↗ |
| Qwen3-235B-A22B | Alibaba | 4,752,000,000,000,000,000,000,000 FLOP | 2025-04-29 | Open weights (unrestricted) | Source ↗ |
| GLM-4.7 | Z.ai (Zhipu AI) | 4,420,000,000,000,000,000,000,000 FLOP | 2025-12-22 | Open weights (unrestricted) | Source ↗ |
Data: Epoch AI · CC BY · View source · Automatically synced; not an exhaustive global sample