Atlas

Models

← All models

Trinity Large Thinking

Arcee AI · 2026-04-01 · 398.6B parameters

Trinity Large Thinking is Arcee AI's reasoning-optimized 398B-total / approximately 13B-active sparse mixture-of-experts model, built on Trinity Large Base and post-trained with extended chain-of-thought instruction tuning and agentic reinforcement learning. Its base model was pretrained from scratch on 17 trillion tokens and then context-extended at a 256K training length for documented 512K-token inference. The checkpoint emits explicit reasoning traces for long-horizon agentic workflows and multi-turn tool use.

Benchmark scores

BenchmarkScore
AA-LCR33.0
AA-Omniscience Index-44.2
Arena Code WebDev — Brand & Marketing1247
Arena Code WebDev — Consumer Product1245
Arena Code WebDev — Content Creation Tools1198
Arena Code WebDev — Data & Analytics1236
Arena Code WebDev — Gaming1256
Arena Code WebDev — HTML1243
Arena Code WebDev — React1233
Arena Code WebDev — Reference-Based Design1235
Arena Code WebDev — Simulations1242
Arena Text — Business, Management & Financial Ops (No Style Control)1331
Arena Text — Chinese (No Style Control)1374
Arena Text — Creative Writing (No Style Control)1309
Arena Text — English (No Style Control)1356
Arena Text — Entertainment, Sports & Media (No Style Control)1299
Arena Text — Exclude Ties (No Style Control)1305
Arena Text — Exclude Ties (Style Controlled)1347
Arena Text — Expert (No Style Control)1359
Arena Text — French (No Style Control)1367
Arena Text — French (Style Controlled)1399
Arena Text — German (No Style Control)1356
Arena Text — Hard Prompts (No Style Control)1342
Arena Text — Hard Prompts English (No Style Control)1355
Arena Text — Instruction Following (No Style Control)1325
Arena Text — Japanese (No Style Control)1310
Arena Text — Korean (No Style Control)1300
Arena Text — Legal & Government (No Style Control)1345
Arena Text — Life, Physical & Social Science (No Style Control)1358
Arena Text — Longer Query (No Style Control)1319
Loading Atlas data…