Thinking Machines Labmodel systemcard

Inkling

#21of 27 · general

xhigh effort collapsed; Inkling Small is a separate model

Standing

dot = this model · shaded = the field's middle 60% · track = full field

General

19.2#21/27

Math

27.4#21/27

Coding

22.0#21/27

Health

19.4#22/26

Engineering

6.2#16/17

Cyber

23.1#11/14

Legal

8.4#25/26

Instruct

32.5#17/27

Agents

29.9#19/27

Professional

22.0#22/27

Multimodal

10.1#17/19

Hallucination

23.9#17/27

Speed

44.5#13/22

Value

28.9#22/27

Open

26.6#9/11

Measured

≈ imputed · amber dot = lab-self-reported among the sources

GPQA Diamond87.1%#21

HLE31.9%#23

MMLU-Pro86.3%#18

LiveBench71.9%#17

SWE-bench V77.6%#18

LiveCodeBench85.5%#14

SciCode46.1%#22

CritPt5.4%#19

Terminal-Bench51.3%#24

GDPval37.0%#17

Arena1443#19

AA Index42.3#19

Vals Index49.3#20

Epoch Capabilities Index149.5#18

Operations

$1 / $4.05$ per 1M tokens, in / out

$1.76blended

77tokens / sec

$0.23$ per solved task

1443arena elo · 7,501 votes

Neighbors

nearest capability fingerprints · deltas describe the neighbor

HY3→ weaker math · cheaper DeepSeek V4 Pro→ stronger professional · cheaper MiniMax M3→ cheaper DeepSeek V4 Flash→ weaker math · cheaper Nemotron 3 Ultra→ cheaper