Inference
EN
Thinking Machines Lab

Inkling

Inkling is a reasoning model from Thinking Machines Lab. 45 benchmarks count toward its index, in 7 categories.

  • Open weights
  • Reasoning
  • Input: text · image · audio
  • Released 2026-07
RankedEnough results for an index.
IndexAll categories, out of 100.54.8±2.9Median 51.0
CoverageIndex weight that has results.95%
SpeedOutput tokens per second.85/sMedian 67/s
First tokenSeconds to the first output token.0.3 sMedian 1.3 s
Input / 1MUS dollars per 1M input tokens.$1.00Median $0.60
Output / 1MUS dollars per 1M output tokens.$4.05Median $2.55
ContextMaximum tokens in one request.524KMedian 262K
EloLMArena rating and rank.1444±4 #69Median 1422

Details

From OpenRouter
API IDThe model ID on OpenRouter.
thinkingmachines/inkling
Max outputMaximum output tokens per request.
472K
Knowledge cutoffLast date of training data.
—
ToolsTool calls on OpenRouter.
Yes
JSON outputOutput follows a given JSON schema.
No
Cache read / 1MUS dollars per 1M cached tokens.
$0.17
WeightsLink to the published weights.
Hugging Face (opens in a new tab) ↗
RetiresDate OpenRouter removes the model.
—

Where to run

APIs that serve this model.
4 providers
ProviderTypeMaker, cloud, host or router.InputOutputContext
Deep InfraDeprecatedHost$0.95$4.05524K
OpenRouterRouter$1.00$4.05524K
Together AIHost$1.00$4.05524K
ModalHost$1.20$5.001.05M

US dollars per 1M tokens, list price. From models.dev (opens in a new tab) (MIT).

Run locally

952.38B parameters
16-bit
2001.0 GB
8-bit
1063.5 GB
4-bit
607.2 GB
Weights
Hugging Face (opens in a new tab) ↗

Memory need with a short context. The 4-bit copy needs more than 128 GB. Compare local models

Capabilities

Index per category, out of 100.
Out of 100
AgenticMulti-step tasks with tools.10 counted51.2121 / 218
CodingCode writing and repair.12 counted54.1106 / 209
ReasoningLogic problems and puzzles.6 counted56.178 / 227
MultimodalTasks with images and text.3 counted55.054 / 84
KnowledgeFacts and expert knowledge.6 counted57.180 / 268
MultilingualTasks in many languages.None counted——
InstructionTasks with strict prompt rules.2 counted66.05 / 63
MathMath problems.6 counted54.168 / 168

Results

45 counted
BenchmarkThe benchmark name.ResultThe score from the source.PlacePlace among results on this benchmark.
AIME 2026Math97.1%5 / 34
LiveCodeBenchCoding85.5% ±1.030 / 142
Instruction Following BenchmarkInstruction79.8%13 / 57
GPQA diamond · Epoch AIKnowledge88.3% ±1.956 / 217
Graduate-Level Google-Proof Q&AKnowledge87.9%39 / 141
Software Engineering Benchmark VerifiedCoding77.6%33 / 114
GPQA Diamond · Vals AIKnowledge87.1% ±1.840 / 135
Artificial Analysis GPQA DiamondKnowledge87.2%59 / 199
OTIS Mock AIME 2024-2025Math88.9% ±3.963 / 200
MMLU ProKnowledge86.3% ±0.343 / 135
Artificial Analysis Humanity's Last ExamKnowledge31.9%72 / 222
Artificial Analysis Omniscience AccuracyKnowledge39.3%56 / 164
Humanity's Last ExamKnowledge46.0%27 / 77
Artificial Analysis Long Context ReasoningReasoning77.3%77 / 217
Critical Physics TasksReasoning5.4%81 / 214
SWE-benchCoding77.6% ±1.934 / 87
ARC-AGI-2 (semi-private)Reasoning36.5%54 / 131
LiveBench Instruction FollowingInstruction70.1%28 / 66
ARC-AGI-1 (semi-private)Reasoning79.5%54 / 126
GDPval-AA normalizedAgentic28.6%65 / 133
MCP AtlasAgentic74.1%26 / 52
CharXiv ReasoningMultimodal82.0%26 / 49
BrowseCompAgentic77.1%34 / 63
GPQA Diamond · BenchLMKnowledge87.9%46 / 84
MMMU ProMultimodal78.5% ±1.051 / 92
Artificial Analysis MMMU-ProMultimodal73.5%63 / 113
Artificial Analysis SciCodeCoding47.0%66 / 114
LiveBench Data AnalysisReasoning72.8%47 / 80
LiveBench MathematicsMath88.4%39 / 66
SimpleQA VerifiedKnowledge40.3% ±1.654 / 91
LiveBench Agentic CodingAgentic49.4%43 / 66
Chess PuzzlesReasoning21.0% ±4.159 / 90
Agent Arena command recoveryAgentic0.133 / 50
SWE-bench ProCoding54.3%57 / 83
Terminal-Bench 2.1 · BenchLMAgentic63.8%61 / 86
τ²-bench BankingAgentic25.0%20 / 28
FrontierMath-Tiers-1-3-v2-PrivateMath33.3% ±2.866 / 90
Humanity's Last Exam without toolsKnowledge30.0%40 / 53
Vibe Code Bench v1.1Coding19.2% ±3.794 / 123
LiveBench ReasoningReasoning78.3%63 / 80
Massive Multi-discipline Multimodal Understanding ProMultimodal73.5%42 / 53
Code MigrationCoding11.8% ±3.463 / 77
CharXiv Reasoning without toolsMultimodal78.1%14 / 17
LiveBench LanguageKnowledge73.5%55 / 66
Terminal-Bench 2.1 · Vals AIAgentic47.6% ±0.463 / 75
LiveBench Coding 2.0Coding64.0%44 / 52
LiveBench ResearchAgentic59.9%44 / 52
LiveBench Browser UseAgentic16.0%42 / 49
FrontierMath-Tier-4-v2-PrivateMath4.9% ±3.462 / 72
LiveBench CodingCoding71.0%57 / 66
LiveBench Long AgenticAgentic14.1%45 / 51
Terminal-Bench 4.0Agentic0.5% ±0.557 / 61
IOICoding14.9% ±5.443 / 45
LiveBench DesignAgentic58.3%50 / 52
Vibe Code Bench 1-100Coding7.3% ±1.825 / 26
Agent Arena task outcomeAgentic−26.852 / 54
SkillsBenchCoding26.1% ±3.934 / 35
ProofBench v1.1Math0.0%53 / 54
Agent Arena steerabilityAgentic−22.353 / 54
CWE-bench v1Agentic37.0%12 / 12
FrontierSWE v2Coding4.1%22 / 22
CorpFin v2Finance68.6% ±0.97 / 131
TaxEval v2Finance75.3% ±0.819 / 143
MedScribeHealthcare85.4% ±1.828 / 109
ARC-AGI-1 (public eval)Reasoning90.6%44 / 116
ARC-AGI-2 (public eval)Reasoning37.8%48 / 122
LegalBenchLegal82.9% ±0.564 / 149
MortgageTaxFinance64.0% ±0.949 / 97
ProgramBenchCoding0.0%34 / 65
MedCodeHealthcare41.2% ±2.256 / 106
CyberBench v1.1Other66.8% ±6.031 / 48
Legal Research BenchLegal28.4% ±3.160 / 90
Finance Agent (v2)Finance46.6% ±0.954 / 79
Public Benefits Bench v1.1Public benefits58.5% ±1.336 / 50
Tax Agent BenchFinance15.3% ±2.461 / 82
EMBFinance38.4% ±3.361 / 75
HLAB-Vals-1.1Legal3.8% ±1.532 / 35
SRE BenchOther0.0%33 / 36

Sources

BenchLMCC BY-NC 4.0OpenRouterTerms restrict reuseEpoch AICC BYLiveBenchNo licence statedVals AINo licence statedARC Prize FoundationTerms restrict reuseSierra ResearchMITLMArenaCC BY 4.0models.devMIT

History

5 releases
Index55.3 → 54.8
23 Oct 202612 Nov 2026
Output / 1M$4.05 → $4.05
23 Oct 202612 Nov 2026

Questions

How much does Inkling cost?
$1.00 per 1M input tokens, $4.05 per 1M output tokens. A free tier is also available.
How good is Inkling?
Its index is 54.8 out of 100. That is place 105 of 283 ranked models.
What is Inkling best at?
Instruction: 66.0 out of 100, place 5 of 63.
How fast is Inkling?
85 output tokens per second, 0.3 s to the first token.
What is the context window of Inkling?
524K tokens.
Is Inkling open source?
Yes. Its weights are published.
When was Inkling released?
July 2026.

Badge

Inkling on Inference360: place 105 of 283, index 54.8
HTML
<a href="https://inference360.ai/models/inkling"><img src="https://inference360.ai/badges/inkling.svg" alt="Inkling on Inference360: place 105 of 283, index 54.8" height="20"></a>
Markdown
[![Inkling on Inference360: place 105 of 283, index 54.8](https://inference360.ai/badges/inkling.svg)](https://inference360.ai/models/inkling)

Compare with

MiMo-V2.554.8Qwen3.6 Plus54.6Gemini 3 Flash55.0GPT-5.4 mini54.5Inkling-Small55.1

Lower price

dots3-note Preview63.4 · FreeApodex 1.1 Mini55.7 · FreeGLM-5.3-Flash63.1 · $0.078/1M

More from Thinking Machines Lab

All models
Inkling-Small55.1Inkling 2 Small66.4