Thinking Machines, the AI lab founded by former OpenAI CTO Mira Murati, released its second model, Inkling Small, on July 31, 2026. The open-weights reasoning model is designed for efficiency, and early benchmark results show it can beat its larger sibling, Inkling, on several demanding tests.
A Smaller Model, a Near-Identical Score
Inkling Small posts a score of 40 on the Intelligence Index, according to independent evaluator Artificial Analysis. Its larger sibling, Inkling, scores 41. That gap is remarkably narrow given the size difference. Inkling Small has 276 billion total parameters but only 12 billion active parameters, which is less than a third of the parameters of Inkling.
Artificial Analysis states that no open model of equal or smaller size scores higher than Inkling Small. That places the new release at the top of its weight class, even as it trails its bigger sibling by a single point on the overall index.
Beating the Bigger Model on Reasoning Tests
On specific benchmarks, the smaller model actually pulls ahead. Inkling Small beats Inkling on Humanity's Last Exam, scoring 32% versus 30%. It also wins on GPQA Diamond, with a score of 89% against Inkling's 87%. Those results suggest the compact architecture handles certain reasoning and coding tasks more effectively than the larger model.
The picture is not entirely one-sided. Inkling Small falls behind Inkling on agent-based tasks and factual knowledge. So the trade-off is real: sharper on some reasoning challenges, weaker on tasks that demand broad knowledge or multi-step tool use.
Token Efficiency That Stands Out
One of the most striking differences is how efficiently Inkling Small uses tokens. It averages 24K output tokens per task. By comparison, Deepseek V4 Flash averages 45K output tokens per task, and GPT-5.4 mini averages 78K output tokens per task. That means Inkling Small produces far less text to reach its answers, which can translate into lower cost and faster responses in production.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
The model handles text, image, and speech inputs, and it comes with a 256K-token context window. That combination makes it a flexible option for developers who need multimodal support without the overhead of a much larger system.
Built for Fine-Tuning
Inkling Small ships under the Apache 2.0 license, and its weights are available on Hugging Face. Users can also fine-tune the model directly in the browser via Tinker Playground. Thinking Machines positions its models as a foundation for fine-tuning with users' own data, and some see fine-tuning as the next frontier in AI.
That strategy makes sense for a lab trying to carve out a niche. Rather than competing purely on raw scale, Thinking Machines is betting that developers want a strong base model they can adapt cheaply and quickly. The browser-based fine-tuning tool lowers the barrier further, letting teams experiment without standing up their own infrastructure.
What the Release Signals
The release of Inkling Small suggests Thinking Machines is willing to ship models that are not just smaller, but in some ways better than their predecessors. The benchmark wins on Humanity's Last Exam and GPQA Diamond are notable, even if the model trails on other dimensions.
For now, the lab's second model is a study in trade-offs. It is more efficient, more open, and more adaptable than its bigger sibling. It also gives up ground on agentic tasks and factual recall. Developers will have to decide which trade-offs matter for their use case.