Trinity Large Thinking
arcee-ai · released 2026-04-01 · other license
Trinity-Large-Thinking is Arcee AI's Apache-2.0 open reasoning MoE (399B total, 13B active, April 2026) for agentic, long-horizon tool use. Artificial Analysis scores it 32; exact per-benchmark numbers are chart-only.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 398.64B total · MoE, — active |
| Architecture | afmoe |
| Context window | 262K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | — |
|---|---|
| SWE-bench Verified | — |
| AIME | — |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | — |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 232.7 GB | 231.2 GB | — | 262K |
Strengths & weaknesses
Strengths: Strong agentic/long-horizon tool use; Above-median open-weights intelligence (AA Index 32) at low cost; Fully open Apache-2.0, 399B total / 13B active, 512K context
Weaknesses: Very verbose; Text-only; per-benchmark numbers published only as charts