GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531
CL-OP46 497
CL-OP48 490 -0.2%
GLM-52 897
GPT-56SC 873
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861
GLM-5 856
MUSE-SPK 841
QWEN-38X 824 -2.3%
GPT-6A 820
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 586 -0.5%
INKL 531
CL-OP46 497
CL-OP48 490 -0.2%
← Back to feed

Inkling Small Scores 40 on AA Index With 12B Active Parameters

Thinking Machines Lab has released Inkling Small, a 276B-parameter open-weight reasoning model with 12B active parameters, text, image, and speech input, and a 256K token context window.

The model’s pitch is not frontier dominance. It is compression. Inkling Small scores 40 on the Artificial Analysis Intelligence Index, one point behind the original Inkling, while using less than a third of the total and active parameters.

The Size Trade

ModelTotal ParametersActive ParametersContextAA Intelligence Index
Inkling Small276B12B256K40
Inkling975B41B1M41

That is the cleanest result in the release. Thinking Machines has moved most of Inkling’s measured intelligence into a much smaller MoE configuration. At the same size tier, DeepSeek V4 Flash max also scores 40 with 284B total and 13B active parameters. MiniMax-M3 reaches 44 with 23B active parameters, while GLM-5.2 max reaches 51 with 40B active.

Inkling Small is not the strongest open-weight model overall. It is a stronger argument that the open-weight middle tier is getting denser: more capability per active parameter, fewer tokens wasted per task, and a smaller serving footprint for teams that cannot afford flagship inference everywhere.

Where Small Beats Flagship

Inkling Small does more than trail the larger model politely. It edges Inkling on several reasoning and coding measures.

BenchmarkInkling SmallInkling
Humanity’s Last Exam32%30%
GPQA Diamond89%87%
CritPt8%5%
SciCode49%46%
Terminal Bench v2.155%55%
GDPval-AA v21269 Elo1237 Elo

The GDPval result is useful because it suggests the smaller model is not simply better at academic benchmark slices. It can also produce outputs that score better on a work-style evaluation, even when the larger model has more raw capacity.

The AA-Briefcase result points in the same direction. Inkling Small reaches 917 Elo against 839 for Inkling, while finishing in 34 turns per task on average instead of 81. The pass rates are almost identical, 20% versus 19%, so the advantage appears to come from presentation quality and task efficiency more than a large correctness gap.

Where It Breaks

The smaller model gives up ground on agentic banking and factual knowledge.

EvaluationInkling SmallInkling
tau3-Banking15%24%
AA-Omniscience Index-92
AA-Omniscience Accuracy31%40%
Hallucination Rate57%63%

The Omniscience result is the warning label. Inkling Small’s hallucination rate is slightly lower than Inkling’s, but its accuracy is materially worse. That means the smaller model is not just making the same number of mistakes with a different style. It is getting fewer facts right.

For production use, that distinction matters. A model can look strong in reasoning and coding while still being the wrong choice for factual research, support knowledge bases, compliance work, or agent tasks where a wrong answer is more expensive than a slow one.

The Token Efficiency Signal

Inkling Small averaged about 24K output tokens per Artificial Analysis Intelligence Index task, slightly below Inkling at about 25K. Peers at similar intelligence levels used far more: DeepSeek V4 Flash max averaged about 45K, and GPT-5.4 mini xhigh averaged about 78K.

That makes Inkling Small interesting even before pricing is fully internalised by the market. Output tokens are where agent costs balloon. A smaller model that writes less, takes fewer turns on AA-Briefcase, and keeps most of the flagship score can be economically more useful than a larger model that wins a narrow benchmark column.

Thinking Machines’ Real Test

Inkling Small gives Thinking Machines a more usable second model. The original Inkling proved the lab could ship a serious open-weight system. Small tests whether it can turn that research result into a deployable product ladder.

The market does not need another model that is vaguely “near frontier.” It needs models that are good enough for a defined workload, cheap enough to run repeatedly, and honest about where they fail. Inkling Small’s profile is unusually legible: efficient reasoning and presentation, weaker factual accuracy, weaker banking-agent execution, and a much smaller active-parameter bill than the flagship.

That is a real product shape. It is not the model you send everywhere. It is the model you route to when the task benefits from reasoning density more than maximum context or factual recall.