What happened
Thinking Machines Lab has released Inkling 2 Small, the latest iteration in its model lineup. The model utilizes a Mixture-of-Experts (MoE) architecture with 594 billion total parameters, of which 17 billion are active per query. It is currently available as a closed-weight product on the company's Tinker platform, with plans to release the weights under an Apache 2.0 license following external safety testing. The model supports image inputs and variable thinking effort, allowing users to adjust reasoning depth at the cost of speed and price.
Thinking Machines Lab launched Inkling 2 Small on November 5, 2026, marking the second generation of its model series. The model is built on a Mixture-of-Experts architecture, which routes requests to a subset of internal 'experts' to optimize compute efficiency. It features 594 billion total parameters with 17 billion active parameters.
The model is currently accessible exclusively through the company's Tinker platform. Pricing is set at $0.30 per 1 million input tokens and $1.00 per 1 million output tokens. It supports multimodal inputs, specifically images, and includes a feature for variable thinking effort, allowing users to scale reasoning depth.
Training for the model involved a large-scale integrated reinforcement-learning process, which the company claims enables the model to utilize visual feedback for tasks such as website and game development. This capability is intended to support agentic workflows where the model performs multistep tasks autonomously.
The company has stated its intent to release the model weights under an Apache 2.0 license, which would allow for commercial use and self-hosting. However, this is contingent upon the completion of external safety testing, and no specific date for this transition was provided.
Source details: cryptobriefing.com
Why it matters
The release of Inkling 2 Small represents a strategic expansion of Thinking Machines' model family, positioning the company to compete in the agentic coding and design sectors. By utilizing an MoE architecture, the model aims to provide high-level reasoning capabilities while maintaining lower compute costs per query. The model's performance on the DesignArena leaderboard—ranking 4th overall and 1st among open-weights models—highlights the potential efficacy of its reinforcement-learning training, which incorporates visual feedback. This release is particularly significant for developers seeking cost-effective, high-performance alternatives for front-end and design-oriented tasks, provided they are willing to operate within the current closed-weight constraints of the Tinker platform.
The model's performance on the DesignArena , where it placed 4th overall and 1st among open-weights models, suggests that the integration of visual feedback during training is yielding tangible results for design-heavy applications.
By offering a model that sits between the original Inkling and Inkling-Small in terms of count, Thinking Machines is providing a middle-ground option for developers who require more reasoning power than the smaller variant but want to avoid the overhead of the larger 975-billion parameter system.
The 'agentic' focus of the model, combined with its ability to handle multistep jobs, aligns with broader industry trends toward autonomous AI agents. The ability to adjust reasoning effort provides a practical mechanism for users to manage the trade-off between latency and accuracy in production environments.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
The primary development to monitor is the transition of Inkling 2 Small from a closed-weight proprietary product to an open-weights model under an Apache 2.0 license. While Thinking Machines has committed to this release, the timeline remains dependent on the completion of external safety testing. Additionally, users should observe how the model's 'variable thinking effort' feature performs in real-world agentic workflows compared to static models. The competitive pricing of $0.30 per million input tokens and $1.00 per million output tokens will also be a key factor in its adoption among teams looking to balance performance with operational costs.
The most critical unknown is the timeline for the Apache 2.0 weight release. Until the weights are released, the model remains a proprietary service, limiting its utility for developers who require local hosting or air-gapped environments.
Market adoption will likely depend on whether the model's performance in agentic coding and design tasks holds up outside of the specific benchmarks cited by the company. Users should monitor independent evaluations of the model's reasoning capabilities in complex, multi-step coding environments.
The long-term viability of the Tinker platform as a primary access point will be tested as the company balances its 'open-weights' branding with the current reality of closed-source distribution.