Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI CTO Mira Murati, has introduced Inkling 2 Mini, a new open-weight AI model designed to deliver advanced reasoning and coding capabilities while being small enough to run on consumer laptops. Announced on November 20, 2026, the model has 26.4 billion total parameters, approximately 3.4 billion active parameters and a memory footprint of around 20GB. Its most unusual feature is a looping architecture that allows selected neural-network layers to run repeatedly, giving the model additional reasoning capacity without substantially increasing its parameter count.
The release is particularly important for the local AI market. Unlike many large AI models that require expensive cloud infrastructure, Inkling 2 Mini is designed to run on local hardware, including laptops, and can operate offline. Thinking Machines says it performs strongly on software-engineering and autonomous-agent benchmarks, positioning it as a competitive option for developers who want to build and run AI agents on their own machines.
What Is Inkling 2 Mini?
Inkling 2 Mini is the latest addition to Thinking Machines Lab’s Inkling family of AI models.
The company describes it as a fast, efficient model optimized for long-horizon reasoning, autonomous coding and tool-based tasks.
Its architecture is based on a Mixture-of-Experts (MoE) design, which allows the model to activate only a portion of its parameters when processing each token.
The model contains:
| Specification | Inkling 2 Mini |
|---|---|
| Total parameters | 26.4 billion |
| Active parameters | 3.4 billion |
| Weight activations per token | 5.4 billion, accounting for looping |
| Architecture | Mixture of Experts with looping |
| Memory footprint | Approximately 20GB |
| Numerical precision | Primarily NVFP4, with final four layers in MXFP8 |
| Availability | Open weights and Tinker platform |
| Main focus | Coding, reasoning and autonomous agents |
The relatively small memory footprint is one of its most important advantages. It makes the model suitable for local deployment on compatible laptops and desktop AI computers instead of depending entirely on remote cloud servers.
What Makes Inkling 2 Mini’s Looping Architecture Different?
The most interesting technical feature is its looping transformer architecture.
Traditional transformer models generally process information through their neural-network layers once before producing an output. Inkling 2 Mini takes a different approach by running selected middle layers multiple times.
The model contains 24 layers divided into four blocks of six layers each. Its two middle blocks are executed three times, while the first and final blocks run once.
This means the model performs more internal computation for each token without needing to store an entirely separate set of parameters for every additional pass.
Thinking Machines believes this approach improves reasoning efficiency, particularly when a task requires several connected steps.
For example, an autonomous coding agent may need to:
- Understand a software project’s existing structure.
- Identify the source of a programming error.
- Examine multiple files.
- Develop a solution.
- Modify the code.
- Run tests and correct failures.
These tasks require more than generating a short answer. They involve maintaining context and making decisions across multiple stages.
The looping architecture is designed to improve the model’s ability to handle this type of extended work.
Inkling 2 Mini Delivers Strong Coding Benchmark Results
Thinking Machines says Inkling 2 Mini performs particularly well on software-engineering benchmarks compared with other open-weight models in its size category.
The company’s published results include:
| Benchmark | Inkling 2 Mini score |
|---|---|
| DeepSWE v1.1 | 48.2% |
| SWE-bench Pro v2 (Hard) | 77.9% |
| SWE-bench Verified | 79.2% |
| SWE-bench Multilingual | 73.7% |
| Vibe Code Bench v1.1 | 77.9% |
| DeepSearchQA | 85.3% |
| BrowseComp | 79.0% |
These benchmarks evaluate different capabilities, including resolving software-engineering problems, building applications, researching information and completing tasks using external tools.
The company says Inkling 2 Mini leads comparable open-weight models under 100 billion parameters on several of these evaluations.
However, these figures come from Thinking Machines’ own published testing. Benchmark results can vary depending on evaluation environments, prompts, software tools and testing configurations.
The model also has limitations. Thinking Machines acknowledges that it is not the strongest option for every type of multi-turn conversation or instruction involving numerous constraints. It also trails larger models on certain knowledge-intensive and advanced reasoning evaluations.
The Model Can Run Offline on a MacBook
One of the biggest selling points of Inkling 2 Mini is its ability to operate locally.
Thinking Machines says the model has an approximately 20GB memory footprint, allowing it to run on compatible consumer laptops and desktop AI systems.
The company demonstrated the model operating on a MacBook Pro with an M5 Pro processor and 48GB of memory while the device was in airplane mode.
During the demonstration, Inkling 2 Mini reportedly built a playable flight simulator within a few minutes.
This is significant because developers can potentially use the model without sending every prompt, file or coding task to an external cloud service.
Local deployment can offer several advantages:
- Privacy: Sensitive files and development projects can remain on the local machine.
- Offline access: The model can operate without an active internet connection once installed.
- Control: Developers can customize and experiment with the model’s weights.
- Reduced cloud dependence: Users can avoid sending every inference request to a paid cloud API.
However, running a model locally still requires compatible hardware and sufficient memory. The approximately 20GB figure refers to the model’s weight footprint, not necessarily the total system memory required for every workload.
Thinking Machines Makes Inkling 2 Mini Open Weight
Unlike proprietary AI systems that can only be accessed through a company’s application or API, Inkling 2 Mini is being released with open weights.
This means developers and researchers can download the model, inspect its parameters, run it locally and experiment with customization.
The model is also available through Thinking Machines’ Tinker platform for inference and customization.
The open-weight approach could encourage developers to build specialized AI agents, coding assistants and research tools using the underlying model.
It also gives the company an opportunity to attract developers who prefer controlling their AI infrastructure instead of relying exclusively on commercial AI services.
AI Safety and the Challenge of Monitoring Internal Reasoning
The looping architecture introduces an important research question: can AI systems become harder to monitor when they perform more internal computation before generating an explanation?
In conventional transformers, reasoning is often expressed through generated text, which researchers can examine to understand how a model approaches a task.
With looping, a model can perform additional internal computations before producing its next token. This means some of its processing may not appear in the written reasoning trace.
Thinking Machines specifically studied this issue while developing Inkling 2 Mini.
The company reports that additional looping improved the model’s ability to complete tasks without generating an explicit chain of thought. Its experiments also found that the written reasoning remained legible in the configurations tested.
However, the company acknowledges that looping increases the amount of computation that can happen without appearing in the model’s visible reasoning. It is releasing related research models to encourage further investigation into monitoring and interpretability.
This makes Inkling 2 Mini relevant not only as a commercial AI model but also as a research platform for understanding how advanced AI architectures work.
Inkling 2 Mini vs Inkling 2 Small
Inkling 2 Mini follows the earlier release of Inkling 2 Small, which Thinking Machines introduced on November 5, 2026.
The two models target different hardware requirements and deployment scenarios.
| Feature | Inkling 2 Mini | Inkling 2 Small |
|---|---|---|
| Total parameters | 26.4 billion | 594 billion |
| Active parameters | 3.4 billion | 17 billion |
| Architecture | Looped MoE | MoE |
| Memory footprint | Approximately 20GB | Requires substantially more memory |
| Primary focus | Local deployment and coding agents | Larger-scale agentic and multimodal work |
| Image input | Supported | Supported |
| Open weights | Yes | Yes |
Inkling 2 Small targets more demanding workloads and offers a much larger overall model capacity. Inkling 2 Mini prioritizes efficiency and accessibility.
The two releases demonstrate Thinking Machines’ strategy of developing models for different deployment environments rather than relying on a single model size.
What This Means for the AI Industry
Inkling 2 Mini arrives as competition in the AI industry increasingly expands beyond large cloud-based models.
Companies are exploring smaller, efficient AI systems that can operate directly on personal computers, smartphones and other devices.
Local AI has several potential advantages, particularly for businesses that handle confidential information or developers who need AI assistance without depending on a continuous internet connection.
The release also demonstrates that model architecture can be as important as parameter count. By repeatedly executing selected layers, Thinking Machines is attempting to extract more reasoning capability from a relatively compact model.
If this approach proves effective across independent evaluations, it could influence how future AI systems are designed.
The Bigger Picture
The launch of Inkling 2 Mini highlights a broader shift towards more efficient and locally deployable artificial intelligence. The industry’s progress is no longer measured only by how many parameters a model contains or how powerful its cloud infrastructure is. Memory requirements, inference speed, energy efficiency, reasoning capabilities and the ability to operate offline are becoming increasingly important.
For Thinking Machines, the release also strengthens its position in the open-weight AI ecosystem. By making a compact model available for developers and researchers, the company is encouraging experimentation with autonomous agents and alternative transformer architectures.
Looking Ahead
The next major test for Inkling 2 Mini will be how it performs outside the company’s demonstrations and benchmark environments. Developers will evaluate its coding reliability, local hardware requirements, inference speed and ability to complete extended tasks. Independent testing will also help establish whether its looping architecture provides consistent advantages over competing models.
For the broader AI industry, the release represents another step towards making advanced AI more accessible beyond expensive cloud infrastructure. If smaller models can deliver increasingly capable reasoning and autonomous coding while operating locally, AI assistants could become a much more integrated part of everyday computers and software development workflows.
SEO Title: Thinking Machines Releases Inkling 2 Mini, an Open AI Model for Laptops
Meta Description: Thinking Machines launches Inkling 2 Mini, a 26.4B-parameter open-weight AI model with a 20GB footprint designed for local coding and autonomous agents.
Suggested URL Slug: thinking-machines-inkling-2-mini-open-ai-model
Focus Keyword: Inkling 2 Mini
Secondary Keywords: Thinking Machines Lab, Mira Murati AI startup, open-weight AI model, local AI models, AI coding model, laptop AI model, looped transformer, autonomous AI agents
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



