Most-used AI models for memory extraction
DeepSeek V3.1 Terminus is the most-used model for memory extraction on OpenRouter, handling 21.1% of requests classified as memory extraction over the last 7 days.
This ranks models by how often they are actually chosen for memory extraction in production traffic, not by benchmark score. Adoption reflects price, latency and availability as much as raw capability — which is exactly why it often disagrees with the benchmark leaderboards, and why it is worth reading alongside them.
| # | Model | Request share | Token share | Input /M | Context |
|---|---|---|---|---|---|
| 1 |  DeepSeek V3.1 Terminus | 21.1% | 0.90% | $0.270 | 164K |
| 2 | DeepSeek V4 Flash 0423 | 11.1% | 13.8% | $0.140 | 1.0M |
| 3 |  MiMo-V2.5 | 8.7% | 16.1% | $0.140 | 1.1M |
| 4 |  Gemini 2.5 Flash Lite | 6.2% | 1.3% | $0.100 | 1.0M |
| 5 |  gpt-oss-20b | 3.6% | 0.70% | $0.030 | 131K |
| 6 |  Step 3.7 Flash | 3.3% | 7.5% | $0.200 | 262K |
| 7 |  Hy3 | 3.2% | 7.8% | $0.132 | 262K |
| 8 |  Qwen3.6 Plus | 3.0% | 0.10% | $0.325 | 1M |
| 9 | nemotron-3-ultra-550b-a55b-20260604:free | 2.6% | 8.3% | — | — |
| 10 | laguna-s-2.1-20260720:free | 2.5% | 5.6% | — | — |
Reading the two columns together: MiMo-V2.5, Step 3.7 Flash, Hy3 take a noticeably larger share of tokens than of requests — meaning they are being used for the longer, heavier memory extraction jobs rather than quick one-shot calls.
Frequently asked
DeepSeek V3.1 Terminus is the most-used model for memory extraction on OpenRouter, handling 21.1% of requests classified as memory extraction over the last 7 days. It is followed by DeepSeek V4 Flash 0423 (11.1%) and MiMo-V2.5 (8.7%).
No. This ranks by how often each model is actually chosen for memory extraction in production traffic through OpenRouter — real-world adoption, which reflects price and availability as much as capability. For capability-based rankings, see the benchmark leaderboards.
Memory Extraction accounts for 1.1% of classified requests and 2.8% of classified tokens on OpenRouter over the trailing 7 days.
Other tasks
Source: OpenRouter (openrouter.ai/rankings), as of 2026-08-04. Shares are of classified, sampled traffic over a trailing 7-day window; absolute volumes are not published.