Model

Trinity-Large-Thinking

OPEN

Arcee AI · Trinity family · released Apr 1, 2026

398B-total/13B-active reasoning MoE pretrained from scratch on 17T tokens by Arcee AI, a small US startup (widely reported as building it 'from scratch to best Meta's Llama') — one of the largest open-weight models produced by an independent US lab. Sibling Trinity Mini (26B/3B active) and Trinity Nano (6B/1B active) were released alongside it but were not independently characterized here. BFCLv4 (70.1) and Tau2-Airline (88.0) scores were reported on the model card but omitted from results below because they don't cleanly match any existing benchmark slug (bfcl-v3 is a different/easier benchmark version, and no tau2-airline slug exists) — flagging bfcl-v4 as a plausible new slug to add rather than misfiling it under bfcl-v3.

ReasoningCodingVisionFunction callingTool useAgentic

1881.7

Elo · rank #69

Parameters

398B

Active params

13B (MoE)

Context

512K tokens

Architecture

Sparse MoE, 256 experts (4 active + 1 shared, ~1.56% sparsity), sigmoid routing, SMEBU load balancing, interleaved local/global gated attention, depth-scaled sandwich norm

License

OpenMDW-1.1

Languages

—

API price (in/out)

$0.25 / $0.8

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

AIMEMath96.3%#9

best: GPT-5.2 · 100.0%

GDPval-AAAgents560#35

best: Claude Fable 5 · 1932

GPQA DiamondReasoning76.3%#96

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning15.0%#73

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning52.3%#35

best: MiniMax M3 · 83.0%

MMLU-ProKnowledge83.4%#36

best: Claude Fable 5 · 91.5%

PinchBenchAgents91.9%#1

best: this model · 91.9%

SWE-bench VerifiedCoding63.2%#72

best: Claude Fable 5 · 95.0%

τ²-Bench AirlineAgents88.0%#1

best: this model · 88.0%

τ²-Bench TelecomAgents94.7%#14

best: Claude Opus 4.6 · 99.3%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Trinity family

Elo progression across releases

Trinity-Large-ThinkingApr 20261881.7

Resources

HFarcee-ai/Trinity-Large-Thinking DOCShttps://arxiv.org/abs/2602.17004 APIAvailable via Arcee AI API

API price $0.25/$0.8 · each benchmark row carries its own source badge (see methodology)