Independent , Honest and Dignified Journalism

Thinking Machines Releases Inkling 2 Mini AI Model Designed to Run on Personal Laptops

The new open weight multimodal model supports text and image processing, with a compact architecture designed to bring advanced AI capabilities to locally operated computers.

SAN FRANCISCO: Artificial intelligence startup Thinking Machines Lab has introduced Inkling 2 Mini, a new open-weight AI model designed to deliver advanced reasoning and coding capabilities while remaining small enough to operate on compatible personal computers.

Released on November 20, the model represents another development in the growing movement towards locally operated artificial intelligence systems that can perform sophisticated tasks without depending entirely on cloud-based computing services.

The company said Inkling 2 Mini uses a specialised architecture that allows it to deliver stronger reasoning performance without requiring the extensive computing resources typically associated with larger AI models.

The model accepts text and image inputs and generates text-based responses, making it suitable for applications such as coding assistants, chatbots, research tools and AI agents.

Released under the Apache 2.0 licence, its open weights allow developers and researchers to examine, customise and integrate the technology into their own applications, subject to the company’s separate acceptable-use policy.

Thinking Machines said the model is designed to run on compatible Apple Silicon computers with at least 32GB of memory, as well as NVIDIA’s DGX Spark desktop AI system.

The release highlights the growing interest in smaller AI models capable of delivering useful performance directly on personal hardware.
A different approach to AI model architecture
Inkling 2 Mini uses a looped transformer architecture, a design that allows selected layers of the model to process information multiple times.

The model contains 24 stored layers, but its middle 12 layers are executed three times for every token.

This creates 48 layer applications per token while retaining the original 24-layer weight structure.

The approach is intended to improve the model’s reasoning capabilities without requiring a proportional increase in the number of stored parameters.

Thinking Machines believes that repeated processing through selected layers can help the model handle tasks that require multiple reasoning steps.

These include complex coding assignments, multi-stage problem-solving and interactions involving AI agents.

The architecture is based on the idea that additional computation within a model can improve performance without necessarily increasing its overall size.

This approach could prove useful as developers seek to build AI systems that are both capable and economical to operate.
Model designed for local computing
One of the most notable features of Inkling 2 Mini is its ability to operate on compatible personal computers.

According to the company, the model has a total of 26.4 billion parameters, with approximately 3.4 billion active parameters in its mixture-of-experts architecture.

Its distributed checkpoint occupies approximately 20GB.

This comparatively compact size makes it possible to operate the model on supported consumer hardware rather than relying exclusively on powerful data centre infrastructure.

Thinking Machines said local inference is supported on Apple Silicon devices with at least 32GB of memory.

The company also supports operation on NVIDIA DGX Spark, a compact desktop AI computing system.

Developers can access the model through several software frameworks, including Ollama, SGLang, vLLM and llama.cpp.

The availability of different deployment options provides flexibility for developers who want to experiment with locally operated AI applications.

Running a model locally can offer several advantages, including greater control over data, reduced dependence on internet connectivity and more predictable access to computing resources.

However, performance will depend on the hardware configuration, memory availability and the complexity of the tasks being performed.
Focus on coding and AI agents
Thinking Machines has positioned Inkling 2 Mini particularly strongly in the area of agentic coding.

AI coding agents are designed to perform more than simply generate individual lines of code. They can examine software projects, understand instructions, make changes across files and work through multi-step development tasks.

Such applications require models capable of maintaining context and handling problems that involve several connected decisions.

The company reported that Inkling 2 Mini achieved a score of 48.2 per cent on the DeepSWE v1.1 benchmark and 46.5 per cent on Vals Index v2.2.

Thinking Machines said the model performed ahead of other open-weight models with fewer than 100 billion total parameters on several evaluations.

These results are based on the company’s reported testing and should be considered alongside independent assessments and real-world performance.

The model is also intended for developers building systems that use external tools and automated workflows.

This makes it relevant to businesses experimenting with software development automation, information retrieval and intelligent digital assistants.
Multimodal capabilities expand its applications
Inkling 2 Mini is not restricted to text-based interactions.

The model can also process images, allowing developers to create applications that combine visual information with language-based reasoning.

Multimodal capabilities are becoming increasingly important because many practical tasks involve photographs, diagrams, documents, screenshots and other visual material.

For example, a coding assistant could potentially examine a screenshot of an application interface and help developers understand its layout or identify visual problems.

Similarly, research applications could use image inputs alongside written instructions to support analysis and information retrieval.

Thinking Machines said the model is designed for general-purpose conversational applications, coding systems and retrieval-augmented generation.

The latter approach combines language models with external information sources, allowing applications to retrieve relevant material before generating responses.

This can be particularly useful when developers need AI systems to work with specialised documents or frequently updated information.
Open-weight release gives developers more flexibility
The decision to release Inkling 2 Mini under the Apache 2.0 licence is another important aspect of the announcement.

Open-weight models provide developers with access to the underlying model parameters, enabling them to deploy the technology on their own infrastructure.

This differs from conventional hosted AI services, where users generally interact with a model through an online platform or application programming interface without directly controlling its underlying weights.

The open-weight approach can provide organisations with greater flexibility in deployment, customisation and experimentation.

Companies handling confidential information may also prefer local infrastructure because it allows them to maintain greater control over where data is processed.

Researchers can examine model behaviour, conduct experiments and investigate how architectural changes affect performance.

Thinking Machines has also released a variable-loop research model alongside Inkling 2 Mini to encourage further investigation into looped transformer architectures.

The company said it hopes researchers will use these releases to explore questions involving reasoning, interpretability and the monitoring of AI systems.
Improving reasoning without increasing model size
A central claim behind Inkling 2 Mini is that repeated processing through selected transformer layers can improve reasoning efficiency.

Traditional transformer models generally move information through their layers in a fixed sequence.

The looped architecture used by Inkling 2 Mini allows certain middle layers to process information repeatedly.

According to Thinking Machines, this design helps the model undertake more internal computation while maintaining a relatively compact stored parameter count.

The company also examined how the architecture affected the model’s ability to solve problems without producing a lengthy written reasoning process.

In its research experiments, a variable-loop version of the model improved its no-chain-of-thought reasoning time horizon from 16 seconds with one pass to 37 seconds with three passes.

The findings suggest that repeated computation can improve performance on tasks requiring sequential reasoning.

However, the company acknowledged that additional looping does not necessarily improve every type of task.

For example, factual recall and knowledge-based questions may not benefit to the same extent as complex reasoning activities.

This distinction is important because model performance depends on more than computational depth alone.

The amount and quality of training data, model architecture and the availability of external tools also influence results.
Privacy and offline operation attract attention
The ability to operate AI models locally has become an important consideration for organisations concerned about data privacy.

Cloud-based AI services generally require information to be transmitted to remote computing infrastructure for processing.

Although service providers offer different privacy and security arrangements, some businesses prefer to keep sensitive information within their own systems.

Locally operated models can help address this requirement by allowing information to be processed directly on compatible devices.

Inkling 2 Mini’s offline capabilities could therefore interest software developers, researchers and organisations seeking more control over their AI workflows.

Thinking Machines described an internal demonstration in which researchers ran the model on a MacBook Pro during a flight without an internet connection.

The demonstration involved creating a playable flight simulator, illustrating the model’s ability to perform coding tasks using local computing resources.

However, offline operation does not automatically guarantee complete security. Organisations still need appropriate device protection, access controls and safeguards against unauthorised data use.
Competition in the smaller AI model market
The release comes amid growing competition among technology companies developing compact AI models.

While major AI laboratories continue to introduce increasingly powerful flagship systems, there is also growing demand for models that can operate efficiently on consumer devices.

Smaller models can reduce infrastructure expenses and make AI technology accessible to developers who do not have access to large computing clusters.

They can also support specialised applications where a relatively narrow set of tasks must be completed quickly and reliably.

Companies are increasingly exploring mixture-of-experts architectures, quantisation techniques and alternative transformer designs to improve performance while reducing memory requirements.

Inkling 2 Mini combines several of these approaches, including sparse expert activation and low-precision computing.

The company said most of the model was trained and deployed using NVIDIA’s NVFP4 numerical format, with the remaining layers using MXFP8.

These techniques reduce the amount of memory required to store and operate the model.
What the release means for developers
The availability of Inkling 2 Mini could expand opportunities for developers building AI applications that do not depend entirely on cloud services.

Software companies could experiment with locally operated coding assistants, while researchers could use the open weights to study the performance of looped transformer models.

Small businesses may also find locally deployed AI systems attractive because they can reduce dependence on recurring cloud computing expenses.

Nevertheless, running a model on a personal computer does not eliminate hardware limitations.

Complex applications may still require powerful processors, sufficient memory and specialised software support.

Developers will also need to evaluate the model’s accuracy, reliability and performance against their specific requirements.

The company has made the model available through its Tinker platform for inference and customisation, while the weights are available through Hugging Face.
A step towards more accessible artificial intelligence
The introduction of Inkling 2 Mini reflects a broader shift in the AI industry towards making advanced capabilities available on smaller and more affordable computing systems.

For several years, the development of powerful AI models has largely depended on large data centres equipped with specialised processors.

However, improvements in model architecture and computing efficiency are creating opportunities to bring more sophisticated AI applications directly to personal computers.

Thinking Machines’ latest release adds another option to this expanding ecosystem.

Its combination of open weights, multimodal capabilities and local deployment support could make it useful for developers interested in building independent AI applications.

The longer-term significance of the model will depend on independent evaluations, practical adoption and how effectively its architecture performs across different real-world tasks.

Nevertheless, the launch highlights an important direction for artificial intelligence development: creating systems that combine strong reasoning capabilities with lower computing requirements and greater user control.

WhatsApp Channel