Model profile / Thinking Machines Lab

Inkling-Small

Inkling Small is an open-weight multimodal mixture-of-experts model from Thinking Machines Lab, with 12B active parameters out of 276B total. It is positioned as the smaller, more efficient member of...

30 Jul 2026Open weights

Sources & details

Four areas of evidence

The capabilities behind the model.

How the indices work

Reasoning

110.9Index points

Mathematical and scientific problem solving.

This model│ Other models · 155 rated

Score and research range

Full-precision point: 110.92807657204804
Range: 106.40–115.46 · 90% Research Range
Scope: R0 latent score conditional on frozen working model and native aggregate portfolio

Score uses the admitted direct native evidence under the selected method.

Coding

No current rating

Software engineering and programming.

Awaiting a qualified category score

General

45.0Index points

Factual reliability and instruction following.

This model│ Other models · 61 rated

Score and research range

Full-precision point: 44.95641579861112
Range: 40.94–48.97 · 90% Research Range
Scope: Fuller G4 after missing instruction quartet I, conditional on recorded R/U

Score includes qualified prediction of missing evidence.

Agentic

98.8Index points

Tool use, workflows and professional tasks.

This model│ Other models · 40 rated

Score and research range

Full-precision point: 98.7843177378487
Range: 96.45–101.12 · 90% Research Range
Scope: Agentic score under joint reference/source reweighting conditional on admitted W/X/P native outcomes

Score uses the admitted direct native evidence under the selected method.

14retained native records

13evidence families

Capabilities estimateQualified original evidence across four areas.

Current index snapshot · 2026-09-16. Each category uses its own scale; scores across categories are not directly comparable. Evidence counts describe the retained Capabilities inputs; a family can contain several tasks or configurations. A point lead does not establish superiority.

Performance in context

Capabilities, price & speed.

Explore every model

Inkling-Small is highlighted wherever a compatible measurement is available.

Inkling-Small: No current checkpoint-bound first-party quote qualified from reviewed original sources; retained values are historical only

Lumina Capabilities Index

Capabilities versus API price

The current Capabilities cohort, with this model highlighted.

Lumina

2025+ · 62 models · 1 highlighted

Chart loads as you explore

Lumina Capabilities Index

Capabilities versus output speed

The current Capabilities cohort, with this model highlighted.

Lumina

2025+ · 72 models · Developer API · 10K · 1 highlighted

Chart loads as you explore

Explore the detail

Sources & model details.

Original evidence, with its context.

Current index evidence 4Scores, research ranges and the sources behind each category.

Capabilities

Score uses the admitted direct native evidence under the selected method.

Exact checkpoint: inkling-small
Method: methods-a4471cee84fd478d2ed3
Range interpretation: C1 native factor under conditional crossed source/family reweighting of admitted fixed-corpus reference and own portfolio
Sources: Artificial Analysis:Artificial Analysis · Epoch AI:Epoch AI · HKUST-NLP:HKUST-NLP · Scale AI:Scale AI
Source cutoff: {"asOfMeaning":"Selected local candidate: retained historical measurements plus reviewed additions; execution dates remain source-native.","newIdentityRecovery":"2026-09-13T08:41:08.804270+00:00","publicProjectionDate":"2026-09-16T06:02:45.491398+00:00","scoring":"Inherited pinned baseline source vintages plus retained4A/4A.1 originals and SWE owner correction; heterogeneous execution dates, often unknown. Not a claim of simultaneous2026-09-13 execution."}

Explore this index

Reasoning

Score uses the admitted direct native evidence under the selected method.

Exact checkpoint: inkling-small
Method: methods-a4471cee84fd478d2ed3
Range interpretation: R0 latent score conditional on frozen working model and native aggregate portfolio
Sources: aa · epoch
Source cutoff: {"asOfMeaning":"Selected local candidate: retained historical measurements plus reviewed additions; execution dates remain source-native.","newIdentityRecovery":"2026-09-13T08:41:08.804270+00:00","publicProjectionDate":"2026-09-16T06:02:45.491398+00:00","scoring":"Inherited pinned baseline source vintages plus retained4A/4A.1 originals and SWE owner correction; heterogeneous execution dates, often unknown. Not a claim of simultaneous2026-09-13 execution."}

Explore this index

Agentic

Score uses the admitted direct native evidence under the selected method.

Exact checkpoint: inkling-small
Method: methods-a4471cee84fd478d2ed3
Range interpretation: Agentic score under joint reference/source reweighting conditional on admitted W/X/P native outcomes
Sources: Artificial Analysis · Scale AI · Toolathlon team (HKUST-NLP)
Source cutoff: {"asOfMeaning":"Selected local candidate: retained historical measurements plus reviewed additions; execution dates remain source-native.","newIdentityRecovery":"2026-09-13T08:41:08.804270+00:00","publicProjectionDate":"2026-09-16T06:02:45.491398+00:00","scoring":"Inherited pinned baseline source vintages plus retained4A/4A.1 originals and SWE owner correction; heterogeneous execution dates, often unknown. Not a claim of simultaneous2026-09-13 execution."}

Explore this index

General

Score includes qualified prediction of missing evidence.

Exact checkpoint: inkling-small
Method: methods-a4471cee84fd478d2ed3
Range interpretation: Fuller G4 after missing instruction quartet I, conditional on recorded R/U
Sources: Artificial Analysis
Source cutoff: {"asOfMeaning":"Selected local candidate: retained historical measurements plus reviewed additions; execution dates remain source-native.","newIdentityRecovery":"2026-09-13T08:41:08.804270+00:00","publicProjectionDate":"2026-09-16T06:02:45.491398+00:00","scoring":"Inherited pinned baseline source vintages plus retained4A/4A.1 originals and SWE owner correction; heterogeneous execution dates, often unknown. Not a claim of simultaneous2026-09-13 execution."}

Explore this index

Benchmark measurements 22Published values in each benchmark’s own units.

22 measurements

Inkling-Small published benchmark measurements
Benchmark	Result	Published configuration	Source
AIME 20262026	95.5%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
ARC-AGI-22025	40.1%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
BrowseComp2025	77.4%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
CharXiv2024	81.3%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
CharXiv Reasoning without tools2024	77.4%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
CritPt2026	8.3%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
Graduate-Level Google-Proof Q&A2023	89.5%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
GPQA Diamond2023	89.5%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
Humanity's Last Exam2025	47.8%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
Humanity's Last Exam without tools2026	31.6%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only

Showing 10 of 22 measurements

Specifications & source recordsDocumented model details and original references.

Model ID / checkpoint: thinkingmachines/inkling-small
Weight access: Open weights
Context window: 524,288
Maximum output: Not documented
Modalities: text, image, audio
Details checked: 1 Aug 2026

BenchLM public datasets — 2026-08-01BenchLM · Published 1 Aug 2026 · checked 1 Aug 2026 · source-checked OpenRouter model catalogue — 2026-08-01OpenRouter · Published 1 Aug 2026 · checked 1 Aug 2026 · source-checked Artificial Analysis LLM data API — 2026-08-01Artificial Analysis · Published 1 Aug 2026 · checked 1 Aug 2026 · source-checked Epoch AI permanent refresh sourceEpoch AI · checked 1 Sept 2026 · source-checked

Specialist task evidenceSeparate agent runs and execution conditions.

All models

Model profile / Thinking Machines Lab

Inkling-Small

30 Jul 2026Open weights

Sources & details

Four areas of evidence

The capabilities behind the model.

How the indices work

Reasoning

110.9Index points

Mathematical and scientific problem solving.

This model│ Other models · 155 rated

Score and research range

Full-precision point: 110.92807657204804
Range: 106.40–115.46 · 90% Research Range
Scope: R0 latent score conditional on frozen working model and native aggregate portfolio

Score uses the admitted direct native evidence under the selected method.

Coding

No current rating

Software engineering and programming.

Awaiting a qualified category score

General

45.0Index points

Factual reliability and instruction following.

This model│ Other models · 61 rated

Score and research range

Full-precision point: 44.95641579861112
Range: 40.94–48.97 · 90% Research Range
Scope: Fuller G4 after missing instruction quartet I, conditional on recorded R/U

Score includes qualified prediction of missing evidence.

Agentic

98.8Index points

Tool use, workflows and professional tasks.

This model│ Other models · 40 rated

Score and research range

Full-precision point: 98.7843177378487
Range: 96.45–101.12 · 90% Research Range
Scope: Agentic score under joint reference/source reweighting conditional on admitted W/X/P native outcomes

Score uses the admitted direct native evidence under the selected method.

14retained native records

13evidence families

Capabilities estimateQualified original evidence across four areas.

Performance in context

Capabilities, price & speed.

Explore every model

Inkling-Small is highlighted wherever a compatible measurement is available.

Inkling-Small: No current checkpoint-bound first-party quote qualified from reviewed original sources; retained values are historical only

Lumina Capabilities Index

Capabilities versus API price

The current Capabilities cohort, with this model highlighted.

Lumina

2025+ · 62 models · 1 highlighted

Chart loads as you explore

Lumina Capabilities Index

Capabilities versus output speed

The current Capabilities cohort, with this model highlighted.

Lumina

2025+ · 72 models · Developer API · 10K · 1 highlighted

Chart loads as you explore

Explore the detail

Sources & model details.

Original evidence, with its context.

Current index evidence 4Scores, research ranges and the sources behind each category.

Capabilities

Score uses the admitted direct native evidence under the selected method.

Exact checkpoint: inkling-small
Method: methods-a4471cee84fd478d2ed3
Range interpretation: C1 native factor under conditional crossed source/family reweighting of admitted fixed-corpus reference and own portfolio
Sources: Artificial Analysis:Artificial Analysis · Epoch AI:Epoch AI · HKUST-NLP:HKUST-NLP · Scale AI:Scale AI
Source cutoff: {"asOfMeaning":"Selected local candidate: retained historical measurements plus reviewed additions; execution dates remain source-native.","newIdentityRecovery":"2026-09-13T08:41:08.804270+00:00","publicProjectionDate":"2026-09-16T06:02:45.491398+00:00","scoring":"Inherited pinned baseline source vintages plus retained4A/4A.1 originals and SWE owner correction; heterogeneous execution dates, often unknown. Not a claim of simultaneous2026-09-13 execution."}

Explore this index

Reasoning

Score uses the admitted direct native evidence under the selected method.

Exact checkpoint: inkling-small
Method: methods-a4471cee84fd478d2ed3
Range interpretation: R0 latent score conditional on frozen working model and native aggregate portfolio
Sources: aa · epoch
Source cutoff: {"asOfMeaning":"Selected local candidate: retained historical measurements plus reviewed additions; execution dates remain source-native.","newIdentityRecovery":"2026-09-13T08:41:08.804270+00:00","publicProjectionDate":"2026-09-16T06:02:45.491398+00:00","scoring":"Inherited pinned baseline source vintages plus retained4A/4A.1 originals and SWE owner correction; heterogeneous execution dates, often unknown. Not a claim of simultaneous2026-09-13 execution."}

Explore this index

Agentic

Score uses the admitted direct native evidence under the selected method.

Exact checkpoint: inkling-small
Method: methods-a4471cee84fd478d2ed3
Range interpretation: Agentic score under joint reference/source reweighting conditional on admitted W/X/P native outcomes
Sources: Artificial Analysis · Scale AI · Toolathlon team (HKUST-NLP)
Source cutoff: {"asOfMeaning":"Selected local candidate: retained historical measurements plus reviewed additions; execution dates remain source-native.","newIdentityRecovery":"2026-09-13T08:41:08.804270+00:00","publicProjectionDate":"2026-09-16T06:02:45.491398+00:00","scoring":"Inherited pinned baseline source vintages plus retained4A/4A.1 originals and SWE owner correction; heterogeneous execution dates, often unknown. Not a claim of simultaneous2026-09-13 execution."}

Explore this index

General

Score includes qualified prediction of missing evidence.

Exact checkpoint: inkling-small
Method: methods-a4471cee84fd478d2ed3
Range interpretation: Fuller G4 after missing instruction quartet I, conditional on recorded R/U
Sources: Artificial Analysis
Source cutoff: {"asOfMeaning":"Selected local candidate: retained historical measurements plus reviewed additions; execution dates remain source-native.","newIdentityRecovery":"2026-09-13T08:41:08.804270+00:00","publicProjectionDate":"2026-09-16T06:02:45.491398+00:00","scoring":"Inherited pinned baseline source vintages plus retained4A/4A.1 originals and SWE owner correction; heterogeneous execution dates, often unknown. Not a claim of simultaneous2026-09-13 execution."}

Explore this index

Benchmark measurements 22Published values in each benchmark’s own units.

22 measurements

Inkling-Small published benchmark measurements
Benchmark	Result	Published configuration	Source
AIME 20262026	95.5%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
ARC-AGI-22025	40.1%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
BrowseComp2025	77.4%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
CharXiv2024	81.3%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
CharXiv Reasoning without tools2024	77.4%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
CritPt2026	8.3%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
Graduate-Level Google-Proof Q&A2023	89.5%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
GPQA Diamond2023	89.5%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
Humanity's Last Exam2025	47.8%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only
Humanity's Last Exam without tools2026	31.6%	Exact BenchLM registry variant Inkling-Small; bulk export does not retain a complete upstream harness configuration.Source-native system	BenchLM public datasets — 2026-08-01 ↗Observed 2026-08-01 · checked 2026-08-01source-checked · reference-only

Showing 10 of 22 measurements

Specifications & source recordsDocumented model details and original references.

Model ID / checkpoint: thinkingmachines/inkling-small
Weight access: Open weights
Context window: 524,288
Maximum output: Not documented
Modalities: text, image, audio
Details checked: 1 Aug 2026

Specialist task evidenceSeparate agent runs and execution conditions.

The capabilities behind the model.

Capabilities, price & speed.

Capabilities versus API price

Capabilities versus output speed

Sources & model details.

Capabilities

Reasoning

Agentic

General

More from Thinking Machines Lab.

Inkling

The capabilities behind the model.

Capabilities, price & speed.

Capabilities versus API price

Capabilities versus output speed

Sources & model details.

Capabilities

Reasoning

Agentic

General

More from Thinking Machines Lab.

Inkling