Key Takeaways

  • Thinking Machines has publicly released the full weights for Inkling, a multimodal Mixture-of-Experts model trained on 45 trillion tokens.
  • The launch signals wider enterprise interest in open weights models for privacy, governance, and domain specialization.
  • Early documentation shows Inkling emphasizing controllable thinking effort, rapid customization, and native multimodal capabilities.

Thinking Machines has introduced Inkling, a large-scale Mixture-of-Experts model whose full weights are now available for download and enterprise deployment. It is a notable moment in a year when open weights models have increasingly outpaced fully open-source models in adoption, a trend highlighted in a Harvard and EPFL working paper from 2025 that documented the drop in training data disclosure and the rise of partial openness. This backdrop matters because enterprises have been pushing for transparency and customizability even as major AI vendors continue to guard their training pipelines.

Inkling arrives with 975 billion total parameters and 41 billion active parameters, paired with a context window of up to 1 million tokens. While the parameter count is high, the model's native multimodal design represents a deeper architectural shift. It processes text, images, and audio natively. That design fits into a broader movement, supported by the Open Source Initiative’s 2024 Open Weights guidance, which suggests that accessible weights increase adaptability even if they stop short of full transparency.

Thinking Machines does not position Inkling as an attempt to outperform frontier proprietary models in raw capability. Instead, the company focuses on controllable thinking effort, broad domain coverage, and a practical experience of fine-tuning on its Tinker platform. This emphasis tracks with analysis from the Oracle research team in a 2024 briefing, which argued that open weights provide appealing cost and privacy advantages for industries that prefer local fine-tuning. Banks, health systems, and government agencies often fall into this category.

A notable demonstration in the release involved Inkling fine-tuning itself using Tinker. The company showed a loop in which the model generated a training dataset, executed a post-training run, evaluated the results, and then switched its own weights—all to become a lipogram model that strictly avoids using the letter "e." It serves as a proof of concept for how far agentic coding and tool use have evolved inside hybrid platform setups, hinting at the potential for domain-specific automation when paired with curated datasets.

Vision and audio have been crowded fields lately, with specialist omni-models attracting attention for their rapid gains. Inkling addresses this natively, integrating capabilities that align with predictions from analysts at the MIT Schwarzman College of Computing, who have repeatedly noted that multimodal reasoning is becoming a baseline expectation for enterprise-grade LLMs rather than a premium feature. If the pattern holds, models that cannot combine text, image, and audio fluently may face operational disadvantages.

The NTIA’s 2024 report on open foundation models, available through the agency’s publication portal, argued that widely accessible weights decentralize power and broaden participation in R&D. Thinking Machines leans into that rationale by making Inkling’s full weights openly available. For organizations that require strict data governance, this provides an immediate on-premises deployment option, though success will depend on internal engineering expertise and risk tolerance.

Thinking Machines also previewed Inkling-Small, a lighter-weight sibling featuring 12 billion active parameters trained using a similar recipe. This release indicates that smaller models continue to challenge assumptions about scale as the sole driver of capability, a trend that IDC analysts have been tracking in ongoing AI workload surveys. Enterprises prioritizing latency or cost per query may find Inkling-Small a more efficient operational fit.

Safety remains a primary consideration for enterprises deploying open weights models. The broader industry conversation, shaped in part by work from groups affiliated with the GAO, continues to emphasize that open weights can be both beneficial and challenging for governance depending on the specific deployment context.

To support enterprise adoption, Thinking Machines has made Inkling available for immediate fine-tuning on its Tinker platform. Developers also gain access to the Inkling Playground within the Tinker console, providing a dedicated interface for direct interaction and qualitative evaluation of the model prior to deployment. These tooling additions point to rapid experimentation and local autonomy as core themes for the release.

Inkling enters a competitive landscape populated by Meta’s LLaMA 3, the Technology Innovation Institute’s Falcon 2, and Mistral AI’s Mistral 7B. While each of these models offers open or widely available weights, Thinking Machines is betting that controllable thinking effort combined with multimodal breadth will differentiate its offering. Whether that proves true will depend on how enterprises balance performance, cost, governance, and the increasingly important question of who controls the underlying training pipeline.