Technology
November 13, 2026 1021 words

Inkling 2 Small: Thinking Machines Lab Releases a 594B Open-Weight Model Under Apache 2.0

Hams AI Tech
Professional Content Team
Expert in AI Prompts & Professional Tools

Thinking Machines Lab, the research company founded by former OpenAI leaders, made its biggest public release yet in November 2026. On November 5, it launched Inkling 2 Small on its Tinker platform with closed weights. One week later, on November 12, after external safety testing, it published the full weights on Hugging Face under the permissive Apache 2.0 license. The model is a 594 billion parameter mixture of experts with 17 billion active parameters, image input, variable effort and a one million token context window.

Why Inkling 2 Small Stands Out

Many large open models come with custom licenses that restrict commercial use. Inkling 2 Small uses Apache 2.0, one of the most business-friendly licenses available. Companies can use, modify and deploy it commercially without special agreements, which makes it an attractive foundation for products and internal tools.

The architecture is also efficient. With only 17 billion parameters active per token, inference costs are closer to a mid-sized model than a giant one, even though the full model holds 594 billion parameters of knowledge.

Key features include:

    • Image input for documents, charts and screenshots.
    • Variable effort, letting users trade speed for deeper reasoning.
    • A one million token context window.
    • A score of 31 on the Artificial Analysis Intelligence Index at launch.

The Safety-First Release Process

Thinking Machines did not publish weights on launch day. It first offered the model through Tinker with closed weights, ran external safety testing and only then released the weights. This staged approach is becoming more common. It gives labs time to identify risks before a model can be freely copied, and it gives customers more confidence in the model's behavior.

A Ready Ecosystem

Inkling 2 Small arrived with broad support from the open-source community:

    • Inference: Unsloth, SGLang, vLLM and TokenSpeed.
    • Post-training: Miles, SkyRL and Snowflake Cortex Training.
    • Hosted APIs: NVIDIA, Fireworks, Together and others.
    • Tinker itself for inference and fine-tuning.

A free tier and a paid tier priced around 0.30 dollars per million input tokens and 1 dollar per million output tokens appeared on model trackers soon after launch.

Fine-Tuning: The Real Opportunity

Tinker is a fine-tuning platform, and that is where Inkling 2 Small becomes especially interesting for professional organizations. An Apache-licensed base model that can be fine-tuned on your own data allows the creation of specialized assistants:

    • A law firm could fine-tune on its precedent library and drafting style.
    • An accounting firm could train on its audit methodology and workpaper formats.
    • An engineering consultancy could teach the model its specification templates and calculation conventions.
    • A healthcare group could adapt it to internal clinical documentation standards, subject to strict privacy controls.

Fine-tuning still requires care: clean data, clear evaluation and a process for keeping confidential information out of training sets where it does not belong. Platforms such as Trajectory, covered in our article on AI developer infrastructure, are emerging to make this easier.

Where Inkling Fits Among Open Models

Inkling 2 Small sits in the middle of a crowded open-weight field. It is much smaller than Kimi K3.1 Pro at 2.8 trillion parameters, but its active parameter count keeps costs low and its Apache license is more permissive. It also served as a teacher model for Nous Research's Arete-27B, showing how quickly open models feed into each other. Our open-weight models roundup compares the main options.

Practical Uses for Professionals

Private Document Analysis

Organizations that cannot send data to external APIs can host Inkling on their own infrastructure or with a trusted provider. Lawyers can use it for contract review and research, combined with structured prompts like the AI legal research assistant workflow.

Internal Knowledge Assistants

A fine-tuned Inkling can answer questions about internal policies and procedures. HR teams can explore human resources prompts for inspiration on the questions employees ask most.

Data and Analytics

Data scientists can use Inkling for code generation and analysis inside secure environments. See Claude Skills for Data Scientists for prompting patterns that translate across models, and browse data science resources.

Considerations Before Deploying

    • Hardware: even with 17 billion active parameters, the full 594 billion parameter model needs substantial GPU memory to serve.
    • Evaluation: test on your own tasks, since a general intelligence score does not predict performance in specialized domains.
    • Governance: document who can access the model, what data it sees and how outputs are reviewed.
    • Updates: open models do not update automatically, so plan how you will adopt future versions.

A Realistic Path for Smaller Firms

Small and mid-sized firms rarely have GPU clusters or machine learning engineers, but they can still benefit from Inkling 2 Small. The practical route is a managed provider that hosts the model in a dedicated environment and signs a data processing agreement. Start by using the hosted model with carefully written prompts for one workflow, such as summarizing client correspondence. If results are strong and volume grows, discuss fine-tuning with the provider using a small, well-curated dataset of your best past work. Measure quality before and after, and keep the original prompts as a fallback. This staged approach lets a firm move from experimentation to a customized assistant without large upfront investment or unnecessary risk to confidential information.

Frequently Asked Questions

Can I use Inkling 2 Small commercially?

Yes. It is released under Apache 2.0, which permits commercial use, modification and redistribution.

Do I need to self-host it?

No. It is available through hosted APIs from NVIDIA, Fireworks, Together and others, as well as on Tinker.

Does it support images?

Yes. Inkling 2 Small accepts image input in addition to text.

What is variable effort?

It lets you choose how much reasoning the model applies, trading speed and cost for depth on harder problems.

The Bottom Line

Inkling 2 Small combines a permissive license, efficient architecture, multimodal input and a strong ecosystem. For organizations that want to own and customize their AI, it is one of the most practical open models available. Explore our agentic AI resources and information technology AI tools to plan your next step.

Related AI News

Topics & Keywords
Inkling 2 Small Thinking Machines Lab Tinker open weights Apache 2.0 mixture of experts