Tools
Compare models
Add up to four models via ?models=slug-a,slug-b — or pick from any model page.
| Spec | DeepSeek |
|---|---|
| Params | 291B / 37B |
| Architecture | moe |
| Context | 1M |
| License | |
| License terms | MIT License |
| 183.4 GBest | |
| Cheapest hosted (out/1M) | $0.084 · OpenRouter |
| Released | Apr 22, 2026 |
| LiveBench | 65.48 |
| LiveBench Agentic Coding | 37.63 |
| LiveBench Coding | 69.23 |
| LiveBench Data Analysis | 68.02 |
| LiveBench Instruction Following | 63.14 |
| LiveBench Language | 70.12 |
| LiveBench Mathematics | 79.65 |
| LiveBench Reasoning | 70.58 |
| Arena Agent | 20th of 55 |
| Arena Agent · Recoverygets back on track after a command fails | 27th of 55 |
| Arena Agent · Steerabilitydoes what it was asked, and changes course when told | 23rd of 55 |
| Arena Agent · Task outcomefinishes what the session set out to do | 18th of 55 |
| Arena Agent · Tool usereaches for the right one, and does not invent one | 25th of 55 |
| Arena Coding | 1484 |
| Arena Creative Writing | 1408 |
| Arena Hard Prompts | 1459 |
| Arena Instruction Following | 1428 |
| Arena Maths | 1425 |
| Arena Text (overall) | 1436 |
| Arena Code (WebDev) | 1581 |
