Inkling Small Scores 40 With 12B Active
Thinking Machines has released Inkling Small, a 276-billion-parameter open-weights reasoning model with 12 billion active parameters.
Artificial Analysis gives it an Intelligence Index score of 40, only one point behind the larger Inkling at 41. It also beats Inkling on Humanity's Last Exam and GPQA Diamond, despite using less than one-third as many total parameters. For developers, the release makes a practical question harder to ignore: how much model capacity is actually needed for a given workload?
At a Glance
• 276 billion total parameters, with 12 billion active during inference.
• Intelligence Index: 40 for Inkling Small versus 41 for Inkling.
• Humanity's Last Exam: 32% versus 30% for the larger model.
• GPQA Diamond: 89% versus 87% for the larger model.
Inkling Small sits near the open-model efficiency frontier.
Inkling Small sits near the open-model efficiency frontier.
A Smaller Route to 40
Thinking Machines, the AI lab led by former OpenAI CTO Mira Murati, has followed its first model with a more compact open-weights release. Open weights means the learned model parameters are available for others to download, inspect, fine-tune, or run under the release terms.
Inkling Small contains 276 billion parameters in total but activates only 12 billion for a given inference step. That sparse design routes work through a limited subset of its capacity. Artificial Analysis scores it at 40, compared with 41 for Inkling, and reports that no open model of equal or smaller total size scores higher.
The Benchmark Trade
The headline result is not simply that the smaller model nearly matches the larger one. Inkling Small actually leads on two demanding evaluations while giving ground elsewhere, which makes the profile more useful than a single composite score.
Measure
Inkling Small
Inkling
Reading
Total parameters
276B
More than 3× larger
Small is under one-third the size
Active parameters
12B
Not stated in the source
Small uses a limited active subset
Intelligence Index
40
41
One-point gap
Humanity's Last Exam
32%
30%
Small leads by 2 points
GPQA Diamond
89%
87%
Small leads by 2 points
Humanity's Last Exam covers difficult questions across many fields, while GPQA Diamond focuses on expert-level science. The source also says Inkling Small trails Inkling on agent tasks and factual knowledge. A model can excel at closed-form reasoning yet struggle with tool calls, state, error recovery, or precise recall.
Efficiency Beyond Model Size
The more interesting claim is token efficiency. According to the source, Inkling Small reaches its answers with substantially fewer generated tokens on average than Inkling. Shorter outputs can reduce response time and inference expense when answer quality holds.
Parameter count, token count, and real cost are separate variables. Routing overhead, memory bandwidth, quantization, batch size, and hardware still shape the final result. For private deployments, 276 billion total parameters are still substantial; open weights improve control but do not erase storage and memory requirements.
What Changes for Builders
The immediate implication is a better candidate for model routing. A team could send code review, difficult science questions, or structured reasoning to Inkling Small while reserving the larger Inkling for workflows that depend on agent reliability or broader factual recall. That is an inference from the published benchmark pattern, not a deployment guarantee.
Before switching, teams should build a small evaluation set from real work. Run both models with the same settings, record tokens and wall-clock time, and judge failed answers separately from merely shorter ones. Open weights also let developers test quantized variants, inspect serving behavior, and keep sensitive prompts inside their own infrastructure.
Limits to Keep in View
These results come through Artificial Analysis and the source article rather than an independent replication presented here. Prompt formatting, inference settings, tool access, and scoring rules can change rankings. The safest reading is that Inkling Small is unusually capable for its size, not that it dominates every comparable model.
The weak spots named in the source should shape adoption tests. A coding agent must choose tools correctly, maintain context, and avoid compounding errors. A factual assistant needs dependable recall or retrieval, even when its answers are concise.
Final Take
Inkling Small narrows the gap with Inkling to one Intelligence Index point while leading it by two points on two reasoning benchmarks. Its 12-billion active footprint and lower token use make efficiency the real story, but agent behavior and factual knowledge remain the checks that matter before deployment. The release strengthens the case for choosing models by workload rather than by size alone.