Inkling

As an AI technology writer, I get to explore a lot of new tools. Most of them offer a slight twist on existing tech, but every now and then, something truly interesting pops up. Today, we're talking about Inkling, a massive 975-billion parameter multimodal AI model, and its partner platform, Tinker, both from the brilliant minds at Thinking Machines Lab. This isn't just another large language model; it's a foundation designed with customization at its core, and it just launched on July 15, 2026.

Let's dive into what makes Inkling and Tinker a game-changer for freelancers, agencies, and anyone serious about building truly custom AI applications.

What is Inkling and What Core Problem Does it Solve?

Inkling is a general-purpose, open-weights multimodal AI model, meaning its underlying code and parameters are freely accessible under the Apache 2.0 license. It was developed by Thinking Machines Lab Inc., a company founded by Mira Murati, formerly the CTO of OpenAI. What sets Inkling apart is its sheer scale and multimodal nature: it boasts 975 billion total parameters (with 41 billion active per token in its Mixture-of-Experts architecture) and can process text, images, and audio inputs to generate text outputs.

The core problem Inkling and Tinker solve is the often-complex and resource-intensive process of customizing large AI models. While many powerful AI models exist, fine-tuning them for specific, niche tasks or proprietary datasets can be a huge headache, requiring deep infrastructure knowledge and significant computational resources. Inkling is explicitly not positioned as the strongest overall model on general benchmarks; instead, it's designed from the ground up to be a highly adaptable base for downstream fine-tuning and specialization. Tinker, the accompanying platform, then takes away the pain of managing the distributed GPU infrastructure needed for this customization, making advanced fine-tuning accessible to a broader range of developers and businesses.

How Does it Work?

At its heart, Inkling is a multimodal autoregressive transformer model. This means it processes information sequentially and predicts the next part of a sequence, much like other large language models. However, its "multimodal" capability allows it to understand and integrate information from different types of data – text, images, and audio – simultaneously. So, you could feed it an image, an audio clip, and some text, and it would generate a coherent text response based on all three inputs.

The "Mixture-of-Experts" (MoE) architecture is a key technical detail. Instead of every part of the model being active for every piece of data, an MoE model routes the input to a subset of specialized "experts." This makes inference more efficient and faster because only a fraction of the total parameters (41 billion out of 975 billion) are active for any given token.

The Tinker platform works by providing a flexible Python API that abstracts away the complexities of distributed training. You write your training loops and define your data, and Tinker handles the heavy lifting of running it across GPUs. It uses techniques like LoRA (Low-Rank Adaptation) for efficient fine-tuning, allowing you to customize Inkling without retraining the entire massive model from scratch.

Key Features

Having explored Inkling and Tinker, here are the features that truly stand out for a freelancer like myself:

• Open Weights & Apache 2.0 License: This is huge. The full weights of Inkling are available on Hugging Face. This means unparalleled transparency, the freedom to inspect and modify the model, and the ability to run it locally if you have the hardware. For a freelancer, this translates to owning your AI stack, reducing vendor lock-in, and building truly bespoke solutions.
• Native Multimodal Reasoning: Inkling isn't just a text model with a tacked-on image encoder. It natively processes text, images, and audio, projecting all modalities into a shared hidden space.
  • Freelancer Use Case: Imagine building an AI assistant for a real estate client that can analyze property photos, transcribe audio notes from a viewing, and read the listing description to generate a detailed, comprehensive property report. Or a content creator's tool that takes video footage, analyzes the visuals and audio, and helps script compelling social media captions.
• Massive Context Window (Up to 1M Tokens): Inkling supports an enormous context window, allowing it to process and remember a vast amount of information. While the Tinker API supports 256K, local deployment can go up to 1M.
  • Freelancer Use Case: Perfect for projects requiring deep understanding of long documents, entire codebases, or extended conversations. Think legal document analysis, comprehensive research assistants, or advanced coding copilots that understand your entire project's context.
• Mixture-of-Experts (MoE) Architecture: This design makes Inkling highly efficient. Despite its 975B total parameters, only 41B are active per token, leading to faster inference compared to dense models of similar scale.
  • Freelancer Use Case: Crucial for applications where speed matters without sacrificing quality, especially when running customized models in production environments for clients.
• Controllable Thinking Effort: Inkling features a "Thinking Effort" dial (from 0.00 to 0.99) that lets you adjust the trade-off between computational cost and output accuracy.
  • Freelancer Use Case: This is fantastic for optimizing costs. For less critical tasks, you can lower the effort for faster, cheaper results. For high-stakes applications, you can crank it up for maximum accuracy.
• Fine-tuning on Tinker Platform: The Tinker platform offers a flexible API for fine-tuning Inkling and other open-source models. It removes the burden of managing distributed GPU infrastructure.
  • Freelancer Use Case: This is where the magic happens for custom solutions. You can fine-tune Inkling on a client's specific data (e.g., customer service logs, product catalogs, internal documentation) to create a highly specialized AI that performs far better than a generic model for their unique needs.
• Agentic and Tool-Use Capabilities: Inkling excels at tasks involving coding, agentic systems, and tool-calling.
  • Freelancer Use Case: Building sophisticated AI agents that can interact with external tools (like APIs, databases, web browsers) to complete complex workflows, automate tasks, or provide intelligent assistance. Imagine an AI agent that can research market trends, summarize findings, and then generate a report, all by itself.

Pricing

The Tinker platform uses a per-million-token pricing model, which is common for AI APIs. This means you pay for what you use, rather than hourly GPU costs. Prices are split across three main meters: prefill, sample, and train.

• Prefill: Processing input/prompt tokens.
• Sample: Generating output tokens.
• Train: Forward and backward pass for gradient computation during fine-tuning.

Some key pricing details:

• Cached Prefill Discount: Input tokens that hit the prompt cache receive an 80% discount.
• Checkpoint Storage: Charged at $0.10 per GB per month.
• Free Credits: New users receive $150 in free credits, which can cover several small experimental runs.

It's important to note that Tinker's prices increased on July 17, 2026, with prefill and sample rates rising by about 50% and train rates by about 10%, attributed to rising compute costs. However, a significant advantage for MoE models like Inkling is that they are priced based on their active parameters, not total size, making them more cost-effective for fine-tuning compared to dense models of similar quality.

While exact current rates for Inkling itself on Tinker aren't explicitly detailed in the provided snippets (an example for Qwen3-8B was given at $0.40/million tokens for training before the price increase), the per-token model with MoE optimization suggests a flexible and potentially efficient pricing structure for fine-tuning complex models.

What Makes it Unique Compared to Similar Tools?

In a crowded AI landscape, Inkling and Tinker carve out a distinct niche:

• Open-Weights & Customization-First Design: Many powerful multimodal models are closed-source and primarily offered as APIs (e.g., GPT-4, Claude). While excellent, they offer limited control over the model's internals or deep customization. Inkling, being open-weights and specifically designed as a fine-tuning base, provides unparalleled flexibility for developers to make the model truly their own.
• Mira Murati's Vision: The backing of Mira Murati and Thinking Machines Lab lends significant credibility and a clear philosophical direction. Their focus on enabling developers to "extend human will and judgment" through customizable AI is a refreshing contrast to the race for the largest, most generalist frontier models.
• Tinker's Managed Fine-tuning: Tinker bridges the gap between raw GPU access (which requires extensive DevOps knowledge) and black-box APIs. It offers Python-level control over training while abstracting away the infrastructure complexity. This "90% algorithmic control, 90% infrastructure headaches removed" approach is highly appealing.
• Controllable Inference Cost/Performance: The "Thinking Effort" dial is a smart feature that directly addresses a common pain point: balancing performance with budget. This granular control is not commonly found in general-purpose API offerings.

Who Should Try This

If you're one of the following, Inkling and Tinker are definitely worth exploring:

• Freelance AI Developers & Consultants: If you build custom AI solutions for clients, Inkling offers a powerful, adaptable foundation. You can fine-tune it to client-specific data, creating unique, high-performing models that generic APIs can't match.
• AI Agencies & Startups: For teams building specialized AI products (e.g., domain-specific chatbots, intelligent agents for automation, multimodal content analysis tools), Tinker provides the infrastructure to efficiently train and deploy custom models without needing a massive in-house MLOps team.
• Researchers & Academics: The open-weights nature and deep customizability make Inkling an excellent tool for cutting-edge AI research, allowing for experimentation with model architectures, training methodologies, and specific use cases.
• Businesses with Proprietary Data: Companies with large, unique datasets (e.g., internal documentation, customer interaction logs, specialized media archives) can leverage Inkling and Tinker to build highly accurate models tailored to their exact business needs, keeping their data private and secure.

Who Should Skip This

Inkling and Tinker might not be the best fit if:

• You Need a Simple, Off-the-Shelf API: If your needs are basic (e.g., simple text generation, summarization) and a generic, widely available API like GPT-4 or Claude suffices, the overhead of fine-tuning and managing a custom model might be unnecessary.
• You Lack Technical Expertise for Fine-tuning: While Tinker simplifies infrastructure, understanding fine-tuning concepts, data preparation, and evaluation metrics is still crucial. This isn't a no-code solution.
• You Have Extremely Limited Compute for Local Runs: While Inkling is open-weights, running the full BF16 weights locally requires significant resources (e.g., 1.9 TB of disk space). Quantized versions are smaller, but still demand considerable hardware. If you only have a consumer-grade laptop and no budget for cloud compute, local experimentation will be challenging.
• You Prioritize Absolute Frontier Performance on General Benchmarks: Thinking Machines Lab openly states Inkling isn't designed to be the "strongest overall model" but rather a balanced, customizable base. If you need the absolute top scores on broad, general-purpose benchmarks out-of-the-box, other models might suit you better.

Final Verdict

Inkling, powered by the Tinker platform, is a significant new entrant in the open-weights AI space. Thinking Machines Lab has delivered a powerful, multimodal foundation model specifically engineered for deep customization and efficient fine-tuning. For freelancers and businesses looking to build specialized, high-performance AI applications on their own data, this combination is incredibly compelling.

The open-weights nature, coupled with Tinker's managed fine-tuning capabilities and the controllable "Thinking Effort," provides a level of flexibility and cost-efficiency that is hard to beat. While it demands a certain level of technical proficiency for fine-tuning, the rewards in terms of tailored performance and intellectual property ownership are substantial.

I give Inkling and Tinker a solid 9/10. It's not for everyone, but for those who need to go beyond off-the-shelf AI and truly make a model their own, it's a stellar offering that promises to unlock a new wave of custom AI innovation.

Frequently Asked Questions

What is Inkling and who developed it?

Inkling is a 975-billion parameter, open-weights multimodal AI model developed by Thinking Machines Lab Inc., a company founded by former OpenAI CTO Mira Murati. It was released on July 15, 2026.

What kind of inputs can Inkling process?

Inkling is a multimodal model that natively accepts text, image, and audio inputs. It generates text outputs based on these combined inputs.

What is the Tinker platform used for?

The Tinker platform, also by Thinking Machines Lab, is a flexible API designed for efficiently fine-tuning open-source models like Inkling. It allows developers to customize models with Python-level control over algorithms and data, while Tinker handles the complex underlying GPU infrastructure.

Is Inkling free to use?

The Inkling model itself is open-weights and licensed under Apache 2.0, meaning its weights are freely available on platforms like Hugging Face. However, using the Tinker platform for fine-tuning and inference has a per-million-token pricing structure, though new users receive $150 in free credits.