Inkling Review: The Open-Weights Model Challenging GPT-5 and Claude

The open-weights model landscape has been moving at a blistering pace throughout 2026. From Google’s Nemotron 3 Ultra to Anthropic’s Claude 4 and OpenAI’s GPT-5, the closed models have been dominating the headlines. But if you’ve been paying attention to the open-source community, you’ve likely noticed a new contender that’s been quietly building momentum: Inkling, Thinking Machines Lab’s first open-weights model, released on July 15, 2026.

Inkling isn’t trying to be the strongest model in every benchmark. Instead, it’s positioning itself as a highly capable, efficient, and customizable base for developers who want open access to a powerful multimodal model. With 975 billion total parameters (41 billion active), a 1 million token context window, and native support for text, images, and audio, Inkling is making a compelling case for open-weights models in the enterprise and developer ecosystems.

In this review, we’ll dive deep into Inkling’s capabilities, pricing, and how it compares to the big players — so you can decide if it’s the right model for your use case.

What Is Inkling?

Inkling is a Mixture-of-Experts (MoE) model from Thinking Machines Lab, released under the Apache 2.0 license. This licensing choice is significant: it means you can use Inkling commercially without worrying about restrictive terms, making it an attractive option for businesses of all sizes.

The model’s architecture is particularly interesting. With 975 billion parameters total but only 41 billion active during inference, Inkling delivers strong performance while keeping compute costs manageable. This MoE design means the model activates only a subset of its parameters for each input, resulting in faster inference times compared to dense models of similar size.

Inkling is multimodal by design, meaning it can process and reason over text, images, and audio natively — no separate models or complex pipelines required. For developers who want to fine-tune the model on their own data, Thinking Machines Lab has made it available on Tinker, their platform for model customization.

How Does Inkling Perform?

Let’s be honest up front: Inkling is not the strongest overall model available today, open or closed. But it’s certainly strong enough to hold its own, and it has some notable achievements.

According to Artificial Analysis, Inkling’s index score of 41 places it ahead of Nemotron 3 Ultra (38) and Gemma, making it one of the stronger open models available. However, it falls slightly behind GLM 5.2, which is widely considered the top open-weight reasoning model in the benchmark set — particularly excelling in pure coding tasks.

For multimodal reasoning, Inkling performs very well. Its ability to process text, images, and audio simultaneously means it can handle complex tasks like analyzing a chart, explaining a diagram, or summarizing a video transcript — all in a single pass. This is particularly valuable for enterprise applications where data comes in multiple formats.

The 1 million token context window is another standout feature. For comparison, most models in the 2026 landscape offer between 128K and 256K tokens. A 1M window means you can feed Inkling entire books, long codebases, or extensive document sets without chunking or summarization. This is a major advantage for applications like legal document review, technical documentation analysis, and long-form content generation.

Inkling vs. The Big Players

How does Inkling stack up against the closed models that have been dominating the conversation? Let’s break it down.

Inkling vs. GPT-5

GPT-5 remains one of the strongest models available, particularly for general-purpose tasks and reasoning. It has the benefit of being backed by OpenAI’s massive infrastructure and continuous updates. However, GPT-5 is a closed model with proprietary weights, meaning you’re locked into OpenAI’s ecosystem and pricing.

Inkling, by contrast, offers open weights and commercial freedom. For teams that want to fine-tune the model on proprietary data or deploy it on their own infrastructure, Inkling’s open-weights approach is a significant advantage. GPT-5 may have a slight edge in raw performance, but Inkling’s efficiency and customization potential make it competitive for many use cases.

Inkling vs. Claude

Claude (Anthropic’s latest generation) is known for its excellent reasoning, long context handling, and strong performance on coding and analytical tasks. Claude’s strengths lie in its clarity and accuracy, particularly for complex reasoning tasks.

Inkling competes well with Claude on multimodal tasks and offers a similar context window advantage. However, Claude’s ecosystem and tooling are more mature, and its pricing is competitive. Inkling’s edge comes in its open-weights flexibility and the ability to run it on your own hardware, which can reduce costs for high-volume users.

Inkling vs. GLM 5.2

GLM 5.2 is currently the top open-weight reasoning model, particularly excelling in coding benchmarks. If your primary use case is code generation and analysis, GLM 5.2 may be the better choice. However, Inkling’s multimodal capabilities give it an edge for applications that need to process images and audio alongside text.

Pricing and Accessibility

One of Inkling’s key selling points is its pricing strategy. As an open-weights model, you have several options:

• Self-hosted: Run Inkling on your own infrastructure with the Apache 2.0 license. This is ideal for enterprises with existing compute resources.
• Cloud API: Access Inkling via Thinking Machines Lab’s API, with pricing that scales based on usage.
• Tinker platform: Fine-tune Inkling on your own data through Tinker, with pricing that depends on the complexity of your fine-tuning needs.

For comparison, GPT-5’s API pricing is competitive but can add up quickly for high-volume users. Claude’s pricing is similarly structured. Inkling’s open-weights model means that once you have the weights, you can use them indefinitely without per-token fees, which can be a significant cost advantage for large-scale deployments.

Pros and Cons of Inkling

Pros

• Open weights: Use commercially under Apache 2.0, no vendor lock-in
• Multimodal: Native support for text, images, and audio
• Large context window: 1 million tokens for long-form content
• Efficient MoE architecture: 41B active parameters out of 975B total
• Fine-tuning ready: Available on Tinker for customization
• Strong open-model rankings: Index score of 41, ahead of Nemotron 3 Ultra

Cons

• Not the strongest overall: Falls behind GLM 5.2 in pure reasoning benchmarks
• Newer model: Less ecosystem maturity compared to GPT-5 and Claude
• Limited third-party support: Fewer integrations and tools available
• Requires more compute: Larger model size means higher infrastructure requirements

Who Should Use Inkling?

Inkling is particularly well-suited for:

1. Enterprises that want to deploy AI models on their own infrastructure and control their data
2. Developers who need multimodal capabilities without managing multiple models
3. Teams that want to fine-tune models on proprietary data
4. Cost-conscious organizations that can benefit from open-weights licensing
5. Applications that process long documents or require large context windows

If you’re looking for the absolute strongest model for a specific task (like coding), you might prefer GLM 5.2. If you want the most mature ecosystem and tooling, GPT-5 or Claude may be better choices. But if you value flexibility, customization, and open licensing, Inkling is a compelling option.

Getting Started with Inkling

Thinking Machines Lab has made it relatively straightforward to get started with Inkling. You can:

1. Download the weights from their GitHub repository
2. Deploy on your infrastructure using their provided scripts
3. Fine-tune on Tinker for custom applications
4. Use the API for quick integration

The model is available on major cloud platforms and can be deployed on-premises, giving you flexibility in how you use it.

FAQ

Is Inkling free to use? Inkling is open-weights under Apache 2.0, so you can use it commercially without licensing fees. However, if you use Thinking Machines Lab’s API or Tinker platform, there are usage-based costs.

How does Inkling compare to GPT-5? GPT-5 is slightly stronger in raw performance, but Inkling offers open weights, multimodal capabilities, and a larger context window. For teams that want to fine-tune or deploy on their own infrastructure, Inkling is competitive.

What is Inkling’s context window? Inkling has a 1 million token context window, which is significantly larger than most models in the 2026 landscape.

Can I fine-tune Inkling? Yes, Inkling is available on Tinker for fine-tuning on your own data.

Is Inkling good for coding? Inkling performs well for coding tasks, though GLM 5.2 currently holds the edge in pure coding benchmarks.

Final Thoughts

Inkling is a strong entry in the open-weights model space, offering a compelling combination of multimodal capabilities, large context window, and efficient MoE architecture. While it’s not the strongest model in every benchmark, its open licensing and customization potential make it an attractive option for enterprises and developers who want flexibility and control.

For teams that are ready to move beyond the closed-model ecosystem and want a model that can grow with their needs, Inkling is worth serious consideration. As Thinking Machines Lab continues to build on Inkling as the foundation of their model family, we can expect to see even more improvements and capabilities in the coming months.

Disclosure: This review contains affiliate links. If you purchase through our links, we may earn a commission at no extra cost to you.