On November 5, 2026, research organization Thinking Machines Lab officially unveiled its new multimodal architecture, designated Inkling 2 Small. The model has been deployed via the Tinker platform, providing technical creators and software developers with direct programmatic access. Built on a sparse architecture, the system contains 594 billion total parameters, with only 17 billion parameters activated during any single inference query. This 17A-594B configuration allows the model to harness the broad representational capacity of massive parameter spaces while maintaining the speed and low compute footprint typically associated with much smaller networks.
The technical efficiency achieved during the training run has drawn immediate scrutiny from across the artificial intelligence sector. According to technical documentation released by Thinking Machines Lab, Inkling 2 Small completed its training cycle in just 68 hours while processing 12 trillion tokens of multimodal data. Achieving stable convergence across 12 trillion tokens in under three days illustrates substantial advances in training pipeline parallelization and cluster throughput. This accomplishment highlights that the computational and temporal barriers previously required to train frontier-grade multimodal models are decreasing rapidly.
Evaluation against established performance standards demonstrated strong competitive capabilities in programmatic media synthesis. In the Code2Video-AA benchmark, which specifically measures the automated generation of moving video sequences driven by programming code, Inkling 2 Small delivered notable results. The model outperformed Meta's competing Muse Spark architecture in this discipline, establishing a new reference point for code-guided video production. This benchmark validates the model's ability to interpret structural coding logic and translate it into coherent, temporally stable visual output.
The model also demonstrated high proficiency when assessed on spatial generation capabilities in three dimensions. In the standardized AA-3D benchmark, which evaluates the fidelity and spatial consistency of generated 3D assets, Inkling 2 Small again surpassed Meta's Muse Spark. Spatial consistency has historically represented a major bottleneck for generative models attempting to interface with video game engines and digital effects pipelines. By generating accurate 3D assets directly from structured inputs, the system offers an alternative to labor-intensive manual asset sculpting.
Deployment through the Tinker ecosystem facilitates the immediate integration of the model into established digital production pipelines across animation, gaming, and interactive media. Performance evaluations place Inkling 2 Small in direct competition with frontier offerings from the largest technology conglomerates. The sparse inference structure addresses one of the primary hurdles in enterprise adoption, namely the prohibitive cost of running multimodal inference at scale. Technical studios can leverage these capabilities to accelerate prototyping and automate repetitive visual asset creation without sustaining massive cloud expenditures.
The release signals intensifying competition in the landscape of code-driven visual synthesis and specialized multimodal modeling. By surpassing Meta's Muse Spark across critical video and 3D benchmarks, Thinking Machines Lab confirms that focused research laboratories can deliver state-of-the-art results outside the dominant platform monopolies. The combination of an abbreviated 68-hour training window, a massive 12-trillion-token training foundation, and an efficient 17-billion active parameter profile establishes a compelling blueprint for next-generation creative tooling.