Thinking Machines Lab has released Inkling, a massive open-weight AI model that represents the strongest American-made foundation model available to the public. The model, which contains 975 billion total parameters with 41 billion active parameters per token, was trained from scratch on 45 trillion tokens of text, images, audio, and video. Unlike many recent AI releases that chase benchmark records, Inkling is positioned as a practical, customizable foundation model designed for real-world applications rather than raw performance maximization.

The release marks a significant milestone for open-source AI development in the United States. Independent evaluators immediately ranked Inkling as the leading U.S.-based open-weight model, ahead of previous benchmarks like Nemotron 3 Ultra and Google's Gemma 4. On the Artificial Analysis Intelligence Index, Inkling debuted at a score of 41, making it the highest-ranking American open-weight release to date. The model also performed competitively on specialized benchmarks: it scored 24% on banking-related tasks, higher than comparable Chinese models, and achieved strong performance on agentic web application tasks, ranking ninth overall in the Design Arena with an Elo rating of 1257, placing it in the same performance band as Claude Opus 4.6 and Google's Gemini 3.5 Flash.

What Makes Inkling Different From Other Large Language Models?

Inkling's architecture incorporates several unusual technical choices that distinguish it from competitors. The model uses a Mixture-of-Experts design, a technique that activates only a subset of the model's parameters for each task, reducing computational cost while maintaining quality. It also employs hybrid attention mechanisms with a 5-to-1 ratio of local to global attention layers, relative positional encoding instead of the more common RoPE (Rotary Position Embedding) approach, and short convolution layers integrated around attention and feed-forward streams. These architectural decisions reflect a focus on efficiency and practical performance rather than pushing benchmark scores to their theoretical limits.

The model supports a context window of up to 1 million tokens, meaning it can process roughly 100,000 words at once. This extended context capability enables the model to handle long documents, extended conversations, and complex multi-step reasoning tasks. Inkling also includes controllable reasoning effort, allowing users to adjust the model's computational intensity based on their specific needs, trading off speed for accuracy when necessary.

How to Deploy and Customize Inkling for Your Use Case

• Immediate Availability: Inkling's full weights are available on Hugging Face under the Apache 2.0 license, with same-day fine-tuning support available through Thinking Machines Lab's Tinker platform and Playground. The model is also integrated with major deployment partners including vLLM, SGLang, Modal, Baseten, and Databricks, enabling developers to deploy it across multiple infrastructure options.
• Multimodal Capabilities: Unlike many multimodal models that suffer performance penalties when processing audio, Inkling maintains strong reasoning quality across text, image, and audio inputs. Users can leverage the model's native multimodal reasoning without sacrificing performance on any single modality, making it suitable for applications requiring cross-modal understanding.
• Smaller Alternative: For teams with tighter computational budgets, Inkling-Small offers 276 billion total parameters with 12 billion active parameters, trained using the same recipe as the larger model. Early evaluations show Inkling-Small achieves competitive performance on several benchmarks despite its significantly lower size and cost, making it practical for cost-sensitive deployments.

The ecosystem support for Inkling's launch was unusually broad. On day one, the model was integrated with quantization tools, community frameworks, and inference optimization libraries, enabling developers to immediately begin experimenting with the model in production environments. This rapid ecosystem adoption suggests strong developer interest in a capable, open-weight American alternative to closed models and Chinese open-source releases.

How Does Inkling Compare to Competitors?

Inkling occupies an interesting position in the AI landscape. While it does not match the absolute peak performance of the most advanced closed models like OpenAI's GPT-4 or the leading Chinese open-weight models such as GLM-5.2 and Kimi K2.6, it represents a substantial step forward for American open-source development. Independent analysts noted that Inkling is "a clear step up from Nemotron Ultra" and represents "the new best American model," though it remains "a bit behind GLM 5.2 on agentic benchmarks and Kimi K2.6 on multimodal tasks".

What distinguishes Inkling from pure benchmark-chasing competitors is its focus on practical reasoning quality. Users and evaluators highlighted the model's "sharp and concise" reasoning output, strong tool-calling capabilities, and good long-horizon error recovery on complex agentic tasks. Rather than generating verbose or rambling responses, Inkling produces focused outputs suitable for real-world applications. Early adopters also noted the model's "quality of mind" and unsycophantic behavior, meaning it avoids excessive agreement with user prompts and maintains independent reasoning.

The model's token efficiency also sets it apart. On the Artificial Analysis Intelligence Index, Inkling averaged 25,000 output tokens per task, significantly lower than GLM-5.2 (43,000 tokens), Kimi K2.6 (38,000 tokens), and DeepSeek v4 Pro (37,000 tokens). This efficiency translates to lower inference costs and faster response times for end users, making Inkling practical for cost-sensitive production deployments.

What Does This Mean for the Open-Source AI Landscape?

Inkling's release signals a maturing open-source AI ecosystem in the United States. The model was developed by a small team that began pretraining in winter 2025 and built coding, reasoning, and agentic capabilities on top from mid-January 2026 onward. This relatively rapid development timeline, combined with the model's competitive performance, suggests that American teams can now compete effectively with well-resourced Chinese AI labs on open-weight releases.

The Apache 2.0 license under which Inkling is released removes legal barriers to commercial use, fine-tuning, and redistribution. This permissive licensing, combined with the model's open weights and broad ecosystem support, positions Inkling as a foundation for downstream applications, research, and specialized fine-tuned variants. Thinking Machines Lab explicitly framed Inkling as a day-one release and foundation for future iterations rather than a final frontier push, suggesting the team plans to continue iterating on the architecture and training approach.

For enterprises and developers seeking to reduce dependence on closed-source AI providers or Chinese-developed models, Inkling offers a credible American alternative with strong practical performance. The model's multimodal capabilities, extended context window, and efficient reasoning make it suitable for a wide range of applications, from customer service automation to complex document analysis and agentic workflows. As the open-source AI ecosystem continues to mature, releases like Inkling demonstrate that American teams can deliver competitive, production-ready models that serve real-world needs.