F2

Speedster · Accelerator

F4
F3

Trinity Large Thinking

Arcee AI

Trinity Large Thinking imagined as a retro AI machineSpeedsterOpen weights
  • Intelligence10.8
  • Hallucination86%
  • Context512K
  • Knowledge2024
  • Speed339tok/s
  • Response8.6s
  • Price$0.39
  • Modalities2
F3 · SpeedsterReasoningSeason 2026
Play with this deck

Trinity Large Thinking

Arcee AI · Built for raw tokens per second. The only cards that can beat the Super Trump.

F3 · Speedstercommon

Card stats

  • Intelligence ▲ higher wins

    Intelligence Index: a composite score from 10 hard evaluations covering reasoning, knowledge, coding, agents and science.

    10.8

    beats 7% of deck

  • Hallucination ▼ lower wins

    Hallucination rate: of the questions a model gets wrong, how often it confidently makes up an answer instead of saying it doesn't know. Lower wins.

    86 %

    beats 13% of deck

  • Context ▲ higher wins

    Context window: how many tokens of input the model can consider at once.

    512K

    beats 29% of deck

  • Knowledge ▲ higher wins

    Training-data cutoff: the most recent month the model learned about. Later wins. N/D means the lab hasn't disclosed it, and an N/D card loses this stat.

    2024

    beats 36% of deck

  • Speed ▲ higher wins

    Median output speed in tokens per second, measured on the labs' own APIs.

    339 tok/s

    beats 94% of deck

  • Response ▼ lower wins

    End-to-end response time: seconds to receive a 500-token answer, thinking time included. Lower wins.

    8.6 s

    beats 90% of deck

  • Price ▼ lower wins

    Blended price per million tokens (3:1 input to output). Lower wins.

    $0.39

    beats 74% of deck

  • Modalities ▲ higher wins

    Number of supported input and output modalities: text, image, audio and video.

    2

    beats 8% of deck

Knowledge: Year only; ranked as January 2024

Model facts

Released
2026-04-01
Weights
Open · Apache 2.0
Parameters
399B (13B active)
Reasoning
Yes
Price in / out
$0.25 / $0.8 per 1M
Time to first token
1.21 s
Humanity’s Last Exam
15.8%
GPQA Diamond
75.2%
SciCode
40.6%

Sources how we source the numbers