Now loading...
Thinking Machines has unveiled Inkling-Small, a new model that delivers performance comparable to its larger predecessor, Inkling, while being significantly smaller and more efficient. The Inkling-Small model is designed as a Mixture-of-Experts transformer with an overall 276 billion parameters, using 12 billion active parameters. It was trained on NVIDIA GB300 NVL72 systems and retains key features such as native reasoning capabilities for audio and images, along with a variable context window of up to 1 million tokens, allowing it to excel across various benchmarks.
When compared to Inkling, Inkling-Small demonstrates similar performance levels, requiring much lower computational resources. It outperforms Inkling in several areas, including agentic tool usage and reasoning tasks. During evaluations across benchmarks such as Terminal-Bench 2.1, HLE text-only, and IFBench, Inkling-Small was noted for its efficiency and remains competitive with other models of similar size.
To enhance usability, the model incorporates variable thinking effort, enabling users to tailor its performance to their specific needs, striking a balance between cost and capability. The complete weights of Inkling-Small are now released and made available for fine-tuning on the Tinker platform as well as for interacting with text, images, and audio.
Inkling-Small underwent several refinements during its training process, building on insights gained from the development of Inkling. Adjustments included optimizing its pre-training data and implementing post-training improvements with an emphasis on reasoning and coding tasks. Results from recent evaluations show that Inkling-Small has outperformed Inkling on various reasoning and agentic benchmarks, although Inkling still retains advantages in knowledge coverage and factuality.
In its multimodal capabilities, Inkling-Small inherits the robust architecture of Inkling, providing excellent functionality for audio intelligence and enhancing its ability to perform visual tasks using Python. With an architecture designed to process audio as dMel spectrograms and images in 40×40-pixel segments, the model achieves substantial versatility, particularly in visual reasoning and longer-form audio reasoning tasks.
Training efforts also focused on epistemics, equipping Inkling-Small with skills in calibration and instruction following. This enables the model to better express confidence levels and make forecasts under uncertain conditions, maintaining a performance level on par with Inkling.
The model’s safety features mirror those of Inkling, incorporating robust safeguards against harmful outputs learned through extensive internal evaluations. Additionally, performance evaluations indicate that Inkling-Small successfully meets industry standards for response safety in sensitive contexts.
With performance enhancements noted across multiple dimensions, including efficiency, reasoning capabilities, and safety, Inkling-Small positions itself as a versatile option for developers seeking excellent return on investment AI solutions. Its launch offers users enhanced functionality across an array of applications, supported by the ongoing commitment of Thinking Machines to develop AI that underscores human decision-making.
For further exploration, interested users can access Inkling-Small and additional models on the Tinker platform, where it is available alongside discounts for early adopters. Comprehensive details, including pricing, can be found on the Tinker pricing page.
