Back to list
Thinking Machines Releases Inkling: A 975B Parameter Open-Weights Mixture-of-Experts Model for Multimodal Customization
Product LaunchOpen SourceLarge Language ModelsMultimodal AI

Thinking Machines Releases Inkling: A 975B Parameter Open-Weights Mixture-of-Experts Model for Multimodal Customization

Thinking Machines has announced the launch of Inkling, a massive open-weights Mixture-of-Experts (MoE) transformer model designed to serve as a flexible foundation for AI customization. Boasting 975 billion total parameters with 41 billion active during inference, Inkling supports a massive 1-million-token context window and was pretrained on a diverse dataset of 45 trillion tokens spanning text, images, audio, and video. Alongside the flagship model, the company introduced Inkling-Small, a more efficient version with 12 billion active parameters. Positioned as a tool to extend human judgment, Inkling emphasizes native multimodal reasoning and controllable thinking effort. It is now available for fine-tuning on the Tinker platform, marking a significant contribution to the open-weights ecosystem.

Hacker News
Share

Key Takeaways

  • Massive Scale and Efficiency: Inkling is a Mixture-of-Experts (MoE) transformer with 975B total parameters, though only 41B are active at any given time to balance performance and cost.
  • Native Multimodality: The model reasons natively across text, images, audio, and video, having been pretrained on a staggering 45 trillion tokens of diverse data.
  • Extensive Context Window: It supports up to 1 million tokens, allowing for the processing of vast amounts of information in a single session.
  • Open-Weights Philosophy: Thinking Machines is releasing the full weights of Inkling to enable users to customize the model for specific needs via the Tinker platform.
  • Model Family Expansion: The release includes a preview of Inkling-Small (12B active parameters), offering a lower-latency alternative for cost-sensitive applications.

In-Depth Analysis

Architecture and Technical Specifications

Inkling represents a significant technical milestone for Thinking Machines, utilizing a Mixture-of-Experts (MoE) architecture. This design allows the model to house a massive 975 billion total parameters while maintaining operational efficiency by only activating 41 billion parameters for any specific task. This approach is central to the model's ability to provide high-level reasoning without the prohibitive computational costs typically associated with dense models of similar scale.

Pretrained on 45 trillion tokens, the model's knowledge base is built from a comprehensive mix of text, images, audio, and video. This broad training allows Inkling to function as a native multimodal system, meaning it does not rely on external plugins to understand different media types but reasons across them inherently. Furthermore, the 1-million-token context window positions Inkling as a powerful tool for long-form content analysis and complex, multi-stage collaborative tasks.

Customization and the Tinker Ecosystem

The primary mission behind Inkling is to build AI that "extends human will and judgment." To achieve this, Thinking Machines has prioritized customization over raw benchmark dominance. While the company acknowledges that Inkling may not be the strongest model currently available in terms of absolute performance, its value lies in its flexibility as an open-weights base.

By making the weights available and integrating the model with the Tinker platform, the developers are encouraging users to "make it their own." The addition of the Inkling Playground within the Tinker console provides a developer-facing interface that allows for immediate interaction and testing. This focus on the "qualitative judgment" of how a model feels during use—rather than just measurable benchmarks—is a core part of the company's strategy to make AI customization accessible for a wider range of use cases.

Balancing Performance with Controllable Thinking

A unique feature of Inkling is its "controllable thinking effort." This allows users to balance the cost of inference with the required performance level, making the model adaptable to various industrial and creative needs. For those requiring even higher efficiency, the introduction of Inkling-Small provides a lighter-weight alternative. With 12 billion active parameters and a training recipe similar to its larger sibling, Inkling-Small is designed to deliver strong performance with significantly lower latency and cost, ensuring that the Inkling family can serve both high-resource and resource-constrained environments.

Industry Impact

The release of Inkling is a notable event for the AI industry, particularly within the open-weights movement. By providing a model with nearly a trillion parameters and native multimodal capabilities, Thinking Machines is challenging the dominance of closed-source providers. The emphasis on a 1-million-token context window and MoE efficiency suggests a shift toward models that are not just large, but also practical for deep integration into human workflows. Furthermore, by focusing on a "balanced foundation" rather than just chasing the top spot on leaderboards, Thinking Machines is highlighting a growing demand for models that are optimized for fine-tuning and specific domain adaptation rather than general-purpose out-of-the-box performance.

Frequently Asked Questions

Question: What makes Inkling different from other open-weights models?

Inkling distinguishes itself through its Mixture-of-Experts architecture (975B total/41B active parameters), a massive 1-million-token context window, and native reasoning across four modalities: text, image, audio, and video. It is specifically designed to be a flexible base for customization rather than just a static foundation model.

Question: How can developers access and customize Inkling?

Developers can access Inkling through the Tinker platform, where it is available for fine-tuning. Thinking Machines has also launched the Inkling Playground in the Tinker console, which provides a dedicated interface for developers to chat with and test the model's capabilities before and during the customization process.

Question: What is Inkling-Small?

Inkling-Small is a lighter-weight version of the flagship model, featuring 12 billion active parameters. It was trained using a similar recipe to the main Inkling model and is intended for users who need lower latency and reduced operational costs while still maintaining strong performance across various tasks.

Read original article

Related News

How to Use LangSmith for Fine-Tuning Open-Source LLMs Like LLaMA2 and GPT-3.5
Product Launch

How to Use LangSmith for Fine-Tuning Open-Source LLMs Like LLaMA2 and GPT-3.5

LangChain has introduced a comprehensive guide detailing how LangSmith supports the fine-tuning and evaluation of Large Language Models (LLMs). The update focuses on enhancing dataset management, providing developers with the tools necessary to refine model performance effectively. The guide specifically highlights practical examples for fine-tuning both open-source models like LLaMA2 and proprietary models such as GPT-3.5. By integrating LangSmith into the fine-tuning workflow, users can better manage datasets and evaluate the outcomes of their training processes. This development marks a significant step in providing structured support for the lifecycle of LLM development, from data preparation to final model evaluation.

Instagram Launches First Draft Feature to Automatically Trim Reels and Highlight Key Video Moments
Product Launch

Instagram Launches First Draft Feature to Automatically Trim Reels and Highlight Key Video Moments

Instagram has introduced a new feature called "First Draft" to its Reels platform, aimed at streamlining the video editing process for creators. The tool automatically trims video clips to focus on the most important highlights, providing a foundational "starting point" for further customization. Currently rolling out to the Instagram iPhone app, First Draft is designed to reduce the manual effort required to edit raw footage into engaging short-form content. By identifying key moments automatically, the feature allows users to quickly transition from capturing footage to the final creative stages of editing. This update reflects Instagram's commitment to lowering the barrier to entry for video creation by offering automated tools that assist in the initial assembly of Reels.

Inside IBM Granite 4.2: A Technical Deep Dive into the New Era of Open-Source Reasoning and Agentic LLMs
Product Launch

Inside IBM Granite 4.2: A Technical Deep Dive into the New Era of Open-Source Reasoning and Agentic LLMs

IBM has officially unveiled Granite 4.2, a groundbreaking family of dense, decoder-only large language models (LLMs) designed specifically for enterprise-grade reasoning and agentic workflows. Released in 3B, 8B, and 30B parameter sizes under the Apache 2.0 license, these models represent a significant leap in open-source AI capabilities. Granite 4.2 is trained on approximately 15 trillion tokens using a sophisticated five-phase strategy that extends its context window to 512K tokens. A key innovation is the introduction of native reasoning—a switchable "thinking" mode that allows the models to perform step-by-step chain-of-thought deliberation. By integrating agentic reinforcement learning (RL) within real-world sandboxed environments like OpenHands and terminal interfaces, IBM has optimized the 8B and 30B versions for complex software engineering and tool-calling tasks, setting a new benchmark for open, transparent, and high-performance AI agents.