Inkling 2 Small

Inkling 2 Small ranks first on LongBench v2.

Thinking Machines Released November 5th, 2026 Family Inkling The lab's model card Compare with others

Benchmarks with a score

Ranked first, out of five or more models

In the top three

Thinking Machines Nov 2026

Inkling 2 Small

Closed weights Mid-size

157.8 #44 of 404

Self-Reported Index, from the lab's own numbers. No Epoch Capabilities Index score yet.

MMMU-Pro 81.6% #11 /90
LongBench v2 70.6% #1 /35
GDP.pdf 40% #14 /19

Where it ranks

Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.

Knowledge and language

Benchmark	Score	Rank	Source
LongBench v2	70.6%	#1 of 35	Lab model card
Humanity's Last Exam	53.2%	#32 of 154	Lab model card
LMArena (Elo)	1,412	#117 of 335	Official leaderboard
AA-LCR 1.1	83.3%	#3 of 6	Lab model card
HLE-Diamond	28.7%	#1 of 2	Lab model card
ComplexConstraints	47.6%	#13 of 15	Official leaderboard
HANDBOOK.md	17.8%	#16 of 16	Official leaderboard
AA-Omniscience v1.1	16.1	#1 of 1	Lab model card

Math and science

Benchmark	Score	Rank	Source
GPQA Diamond	93%	#13 of 302	Lab model card
HMMT 2026	95.8%	#2 of 19	Lab model card
ARC-AGI	96.5%	#16 of 95	Official leaderboard
ARC-AGI-2	81.7%	#20 of 98	Official leaderboard
MathArena (Apex, IMO, USAMO)	63.5%	#4 of 12	Lab model card
BioMysteryBench	70%	#15 of 30	Official leaderboard
MysteryMechanism	25.2%	#19 of 31	Official leaderboard
Terminal-Bench-Science	5.70%	#12 of 12	Lab model card
CritPt v2	55.7%	#1 of 1	Lab model card

Coding

Benchmark	Score	Rank	Source
DesignArena Game Dev	1,433	#2 of 143	Official leaderboard
DesignArena Code	1,384	#4 of 143	Official leaderboard
DesignArena 3D Design	1,460	#4 of 136	Official leaderboard
DesignArena Website	1,332	#6 of 151	Official leaderboard
SWE-bench Verified	82.2%	#7 of 155	Lab model card
DesignArena Frontend Web App	1,286	#5 of 53	Official leaderboard
DesignArena Agentic Overall	1,269	#11 of 74	Official leaderboard
Code Arena	1,646	#20 of 128	Official leaderboard
SWE-bench Multilingual	81.2%	#12 of 73	Lab model card
Vibe Code Bench	81.5%	#30 of 120	Official leaderboard
SWE-Bench Pro V2	97.7%	#5 of 18	Official leaderboard
DeepSWE	72.1%	#16 of 56	Lab model card
SWE Atlas v1.1	61.9%	#13 of 28	Official leaderboard
SWE-Bench Pro V2 Hard	86.3%	#9 of 18	Official leaderboard
SWE-bench Multimodal	43.1%	#7 of 12	Lab model card
FrontierCode 2.0	22.6%	#20 of 32	Official leaderboard
sudo L7	14.1%	#12 of 18	Official leaderboard
SWE-Marathon v1.1	38.8%	#17 of 20	Official leaderboard
FrontierCS 2	8.33%	#19 of 19	Official leaderboard
PostTrainBench 1.2	31.9%	#3 of 3	Lab model card
AA-3D	1,021	#1 of 1	Lab model card
Code2Video-AA	1,020	#1 of 1	Lab model card
PMPP-Hard	494.3	#1 of 1	Lab model card

Agents and tools

Benchmark	Score	Rank	Source
BrowseComp	90.2%	#9 of 81	Lab model card
MCP Atlas	85.4%	#6 of 37	Official leaderboard
τ³-bench Banking	40.4%	#4 of 22	Lab model card
ScreenSpot-Pro	83.8%	#7 of 32	Lab model card
Agent Arena	2.51%	#18 of 56	Official leaderboard
EnterpriseOps-Gym	46.8%	#1 of 3	Lab model card
DeepSearchQA	90.9%	#8 of 23	Lab model card
AutomationBench	45.6%	#17 of 47	Lab model card
Terminal-Bench 4.1	13%	#33 of 43	Official leaderboard

Professional work

Benchmark	Score	Rank	Source
EnterpriseBench: CoreCraft	71.8%	#3 of 51	Official leaderboard
Vals Finance Agent v2.1	67.5%	#5 of 40	Official leaderboard
Tax Agent Bench	38.9%	#13 of 84	Official leaderboard
Harvey LAB v1.1 (HLAB-Vals-1.1)	21.7%	#6 of 38	Official leaderboard
JobBench	70.5%	#3 of 15	Lab model card
Legal Research Bench	46.6%	#19 of 93	Official leaderboard
Vals Legal Index	73.9%	#8 of 38	Official leaderboard
Vals Finance Index	61.5%	#9 of 40	Official leaderboard
DesignArena Python-PPTX Slides	1,210	#10 of 36	Official leaderboard
DesignArena Spreadsheet	1,262	#10 of 35	Official leaderboard
APEX-Accounting	62%	#17 of 49	Official leaderboard
DAYJOB: Finance	14.3%	#14 of 39	Official leaderboard
Tuesday Work Index	54.6	#17 of 44	Official leaderboard
Excel Modeling Benchmark v1.1	69.5%	#21 of 42	Official leaderboard
Vals Index v2.2	56.1%	#22 of 37	Official leaderboard
DeepFinance	29.8%	#15 of 23	Official leaderboard
APEX-Agents	46.8%	#41 of 61	Collected by Epoch AI, source not named
GDPval-AA v2.1	1,630	#7 of 10	Lab model card
APEX-Agents 1.1	60.9%	#47 of 65	Official leaderboard
AA-Briefcase v1.1	1,527	#7 of 8	Lab model card
DiligenceBench	58.1%	#1 of 1	Lab model card

Vision, audio and robotics

Benchmark	Score	Rank	Source
MMMU-Pro	81.6%	#11 of 90	Lab model card
MATH-Vision (MathVision)	92.8%	#8 of 51	Lab model card
CharXiv	85.6%	#19 of 75	Lab model card
Humanity's Sixth Sense	33.3%	#13 of 34	Official leaderboard
BabyVision	68.6%	#13 of 26	Lab model card
PerceptionBench (Moonshot)	64.7%	#3 of 5	Lab model card
GDP.pdf	40%	#14 of 19	Official leaderboard
Chartography	36.7%	#13 of 14	Lab model card

Safety and security

Benchmark	Score	Rank	Source
StrongREJECT	97.8%	#4 of 19	Lab model card
CyberBench v1.1 (Vals)	72.7%	#15 of 44	Official leaderboard
SRE-Bench	1.91%	#25 of 35	Official leaderboard
Apollo Research strategic deception evaluations	8.00%	#3 of 3	Lab model card
Covert misbehavior (Thinking Machines)	11.8%	#1 of 1	Lab model card
Destructive actions (Thinking Machines)	4.70%	#1 of 1	Lab model card
DistressBench	82%	#1 of 1	Lab model card
HoneyBench v0.2-alpha	44.0	#1 of 1	Lab model card
Misbehavior (Thinking Machines)	14.9%	#1 of 1	Lab model card
PatchEval	79.6%	#1 of 1	Lab model card

Other

Benchmark	Score	Rank	Source
SAGE v1.1 (Vals)	88.6%	#8 of 29	Official leaderboard
Vals Science Index	32.7%	#17 of 25	Official leaderboard
Artificial Analysis Intelligence Index v5	30.6	#2 of 2	Lab model card

Its lab's numbers put it at 157.8, number 44 of 404

The Self-Reported Index fits one score to the 24 capability results Thinking Machines reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 154.8 to 160.8. How we compute it.

Its lab's numbers next to 28 leaderboards

Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.

Benchmark	Lab says	Board says	Gap
APEX-Accounting	14.5% Apex-Accounting v1.1 pass@1 (mean score 62.0)	62% Leaderboard, loop_agent	−48 pts
Tax Agent Bench	65.8% Tax Agent Bench, 65.8%	38.9% Leaderboard	+27 pts
GDP.pdf	25.2% GDP.pdf	40% Leaderboard, xhigh	−15 pts
DesignArena 3D Design	1,401 DesignArena 3D Design	1,460 Leaderboard	−59
ComplexConstraints	44.5% ComplexConstraints v1.0	47.6% Leaderboard, xhigh	−3.1 pts
FrontierCode 2.0	25.4% FrontierCode v2.0	22.6% Leaderboard, max, Devin CLI	+2.8 pts
DesignArena Frontend Web App	1,296 DesignArena Web Apps	1,286 Leaderboard	+10
DesignArena Website	1,340 DesignArena Website	1,332 Leaderboard	+8
Code Arena	1,653 Code Arena, 1653 Elo	1,646 Leaderboard	+7
MCP Atlas	85% MCP Atlas (Public; 85.4 on Full)	85.4% Leaderboard	−0.4 pts
DesignArena Game Dev	1,428 DesignArena Game Dev	1,433 Leaderboard	−5
Vibe Code Bench	81.6% Vibe Code Bench v1.1	81.5% Leaderboard, OpenHands	Same
Vals Finance Agent v2.1	67.5% Finance Agent v2.1, 67.5%	67.5% Leaderboard	Same
CyberBench v1.1 (Vals)	72.7% CyberBench v1.1	72.7% Leaderboard	Same
Legal Research Bench	46.6% Legal Research Bench, 46.6%	46.6% Leaderboard	Same
SWE-Bench Pro V2 Hard	86.3% SWEBench Pro v2 (Hard)	86.3% Leaderboard	Same
Vals Finance Index	61.5% Vals Finance Index	61.5% Leaderboard	Same
SAGE v1.1 (Vals)	88.6% SAGE v1.1, 88.6%	88.6% Leaderboard	Same
Harvey LAB v1.1 (HLAB-Vals-1.1)	21.7% Harvey LAB v1.1, 21.7%	21.7% Leaderboard	Same
Vals Index v2.2	56.1% Vals Index (v2.2)	56.1% Leaderboard	Same
Vals Legal Index	73.9% Vals Legal Index	73.9% Leaderboard	Same
BioMysteryBench	70% BioMysteryBench	70% Leaderboard, Vals AI harness, v11 full set	Same
DAYJOB: Finance	14.3% DAYJOB (Finance)	14.3% Leaderboard, xhigh	Same
EnterpriseBench: CoreCraft	71.8% EnterpriseBench-CoreCraft	71.8% Leaderboard, xhigh	Same
Humanity's Sixth Sense	33.3% HSS	33.3% Leaderboard	Same
sudo L7	14.1% Sudo L7, 14.1%	14.1% Leaderboard, xhigh	Same
SWE-Marathon v1.1	38.8% SWE-Marathon v1.1, 38.8%	38.8% Leaderboard, xhigh, Claude Code	Same
Excel Modeling Benchmark v1.1	69.5% Excel Modeling Benchmark v1.1, 69.5%	69.5% Leaderboard	Same

Its launch post

The benchmarks the lab chose to headline, with the scores it gave.

Introducing Inkling 2 Small (Thinking Machines) Nov 5, 2026

Vals Index (v2.2) 56.1
FrontierCode v2.0 25.4
DeepSWE v1.1 72.1
Terminal Bench Science v0.1 5.7
Vibe Code Bench v1.1 81.6
SWEBench Pro v2 (Hard) 86.3
PostTrainBench v1.2 31.9
BioMysteryBench 70
MCP Atlas (Public; 85.4 on Full) 85
JobBench 70.5
EnterpriseOps-Gym 46.8
EnterpriseBench-CoreCraft 71.8
Vals Finance Index 61.5
Tau 3 Banking v1.0.1 40.4
DAYJOB (Finance) 14.3
Apex-Accounting v1.1 pass@1 (mean score 62.0) 14.5
Vals Legal Index 73.9
DesignArena Web Apps 1296
DesignArena Website 1340
DesignArena Game Dev 1428
DesignArena 3D Design 1401
PerceptionBench 64.7
ScreenSpot Pro (original; 91.5 with Python) 83.8
HSS 33.3
BabyVision (original; 93.7 with Python) 68.6
GDP.pdf 25.2
MMMU Pro (Standard 10) 81.6
Charxiv RQ (original; 87.7 with Python) 85.6
Chartography (original; 80.4 with Python) 36.7
CritPt v2 (Full; 35.0 on Hard) 55.7
HLE-Diamond 28.7
HLE (with tools) 53.2
Apex 2025 63.5
LongBench v2 70.6
ComplexConstraints v1.0 44.5
DistressBench 82
FORTRESS (adversarial; benign 79.7)
StrongREJECT (none) 97.8
CyberBench v1.1 72.7
PatchEval 79.6
AA Index (v5.0) 30.6
AA-Briefcase v1.1 1527
GDPVal-AA v2.1 1630
AA-LCR v1.1 83.3
AA-Omniscience v1.1 (non-hallucination rate 73.3%) 16.1
AA-3D 1021
Sudo L7, 14.1% 14.1
SWE-Marathon v1.1, 38.8% 38.8
SWE-bench Multilingual, 81.2% 81.2
SWE-bench Multimodal v2, 43.1% 43.1
SWE-bench Verified, 82.2% 82.2
BrowseComp with context management, 90.2% 90.2
DeepSearchQA, 90.9% 90.9
AutomationBench v1.0.6, 45.6% 45.6
Finance Agent v2.1, 67.5% 67.5
Excel Modeling Benchmark v1.1, 69.5% 69.5
Tax Agent Bench, 65.8% 65.8
Harvey LAB v1.1, 21.7% 21.7
Legal Research Bench, 46.6% 46.6
Code Arena, 1653 Elo 1653
MathVision, 92.8% 92.8
SAGE v1.1, 88.6% 88.6
GPQA Diamond, 93.0% 93
HMMT Feb 2026, 95.8% 95.8
HoneyBench v0.2-alpha Reward Hacking, 44, lower is better 44
PMPP-Hard AUC, 494.27 task-hours 494.3
Misbehavior (Internal), lower is better, alignment table 14.9
Covert misbehavior (Internal), lower is better, alignment table 11.8
Destructive actions (Internal), lower is better, alignment table 4.7
Covert Action with No Goal (Apollo Research) 8, Covert Action with Goal 0, lower is better, alignment table 8
Diligence Bench, best harness, as reported by Thoughtful 58.1
Code2Video-AA (Artificial Analysis, quoted in the launch post) 1020