‹ PublicAI Index
The LLM benchmark aggregator.
Inkling
Thinking Machines · 952B · open weights
Strongest in Jailbreak resistance (#2 of 268), weakest in Agents (#273 of 298). Above par in 20 of 39 scopes. Among the models it meets almost everywhere, it finishes behind Claude Fable 5.1 and Kimi K3 and ahead of DeepSeek V4 Flash and MiniMax M3.
§ 1 · Profile
What it is good at
Bars run from 50 — the average of the models each source lists — so right of the line is above par. The middle column is the gap to whoever leads that scope.
Safety59.9−1.5#5/3291/3
Jailbreak resistance64.4−0.6#2/2681/1
Secure code65.2−1#10/2701/1
Harm refusal58.6−2.9#20/2921/2
Fairness60.9−9.1#38/2921/2
Toxicity avoidance56.1−2.9#70/2681/1
Knowledge52.6−9.9#73/1971/3
Academic knowledge56.3−7.6#42/1231/1
Factuality55.6−5.2—/0✱✱
Human preference60.8−8#76/3541/1
Human preference60.8−8#76/3541/1
Professional49.6−12.8#103/1942/2
Cybersecurity51.7−12.2#30/471/1
Document design26−42.7#31/311/1
Finance55.8−6.6#39/1621/1
Medical52.5−14#72/1491/1
Legal46.5−19.4#115/1731/1
Reasoning48.2−17.2#110/1955/5
Science57.7−5.8#40/1221/1
Mathematics48.9−14.9#92/1521/2
Reasoning44.3−24.7#93/1473/3
Expert reasoning53.4−13—/0✱✱
Coding42.3−23.9#151/1793/6
Agentic coding42−24.9#146/1733/5
Code generation58.2−7.1—/20✱0/1
Repository Q&A36−30.7—/0✱✱
Core abilities44.3−21#154/1993/4
Data analysis49.2−16#21/311/1
Instruction following53.7−20.1#28/621/1
Language44−22.6#51/621/1
General intelligence37.3−32.1#167/1782/3
Long context56−8.6—/0✱✱
Agents41.1−25.3#273/2983/6
Long-horizon work34.9−39.1#30/311/1
Web research35.6−28.3#33/351/2
Computer use33.4−34.7#65/681/2
Knowledge work51.3−19.4#88/2082/2
Tool use46.2−27.8—/81✱0/1
Workflow automation44−22—/0✱✱
§ 2 · Head to head
What it beats, and what beats it
The same models turn up scope after scope. Counted once: where both were placed, who finished higher. Bars run right for Inkling, left for the other.
§ 3 · Sources
Where the numbers come from
11 publications, 61 figures. Every one links to the page it was read from.
LMArena Text 1443
Artificial Analysis Intelligence Index 12
GDPval-AA 1080
AA-Briefcase 833
ARC-AGI-2 36.5%
LiveBench 44.3LiveBench · Agentic 9.8LiveBench · Reasoning 78.3LiveBench · Coding 48.7LiveBench · Browser Use 12LiveBench · Data Analysis 72.8LiveBench · Research 60.3LiveBench · Design 55.2
LiveBench 1.0 71.9LiveBench 1.0 · Reasoning 78.3LiveBench 1.0 · Coding 71LiveBench 1.0 · Agentic Coding 49.4LiveBench 1.0 · Mathematics 88.4LiveBench 1.0 · Data Analysis 72.8LiveBench 1.0 · Language 73.5LiveBench 1.0 · Instruction Following 70.1
SimpleBench 50%
Vals · Legal Research Bench 28.36%Vals · LegalBench 82.95%Vals · Harvey Legal Agent Benchmark 3.75%Vals · Finance Agent 46.6%Vals · CorpFin 68.57%Vals · TaxEval 75.31%Vals · MortgageTax 63.99%Vals · MedCode 41.19%Vals · MedScribe 85.41%Vals · CyberBench 66.79%Vals · SWE-bench Verified 77.6% (Mini-SWE-agent)Vals · Vibe Code Bench 19.21% (OpenHands)Vals · Code Migration 11.79%Vals · GPQA Diamond 87.12%Vals · MMLU Pro 86.3%
Enkrypt · Jailbreak risk 0.7%Enkrypt · Harmful content risk 2.2%Enkrypt · CBRN risk 8.7%Enkrypt · Toxicity risk 2%Enkrypt · Bias risk 61%Enkrypt · Insecure code risk 2.2%
GDPVal-AA 1234tau3-Banking 29.1%Toolathlon Verified 45.5%Automation Bench Public 12.8%Apex-Agents (pass@1) 19%MCPMark 51.2%BrowseComp 77.1%WildClawBench 52.3%Terminal-Bench 2.1 55.1%SciCode 46.1%SWE-Atlas-QnA 25.5%SWE Bench Pro 43.1%Humanity's Last Exam (without tools) 31.9%GPQA Diamond 87.2%CritPt 5.4%AA-LCR 73.3%AA-Omniscience Accuracy 42%AA-Omniscience Non-Hallucination 32%
Badge
[](https://publicai.io/model-index/m/inkling)