Read this article in: Portuguese (Português) · Spanish (Español)
Artificial Intelligence (AI) is undergoing an accessibility revolution, and it is in this scenario that Thinking Machines Lab just unveiled the Inkling 2 Mini. The model, announced on November 20, represents a notable advancement because it is an open-source model that can operate completely offline on consumer computers, such as Apple MacBooks. With only 26 billion parameters, the Inkling 2 Mini challenges models of gigantic scale, proving that it is possible to deliver robust performance on complex tasks, such as reasoning and programming, without depending on expensive cloud infrastructure.
Inkling 2 Mini: High-Level Performance on Personal Hardware
The launch of the Inkling 2 Mini is seen as an extremely efficient alternative compared to massive models, such as the Inkling 2 Small, which has 594 billion parameters. The model’s major highlight is its ability to run locally. To achieve this, it has a memory footprint of approximately 20GB, which allows it to run on Apple Silicon machines with at least 32GB of unified memory, utilizing platforms like Ollama.
In addition to portability, the Inkling 2 Mini’s performance is highly competitive. It achieved an Artificial Analysis (AA) intelligence index of 20.1. Test results show that the model outperforms significantly larger competitors in critical software engineering metrics. For example, in DeepSWE v1.1 tests, it achieved 48.2%, surpassing the 42.2% recorded by Qwen3.8-27B, and reached 77.9% on the SWE-bench Pro v2 Hard, an index superior to the 73.8% of the compared model.
The model also demonstrates versatility, accepting text and image inputs and featuring a context window that can extend up to one million tokens. To demonstrate its power in the field, the company reported that the Inkling 2 Mini operated completely offline on a MacBook Pro M5 Pro with 48GB of memory, successfully generating a playable flight simulator in minutes during a flight.
📊 Performance Comparison in Benchmarks
| Benchmark | Inkling 2 Mini | Competitor |
|---|---|---|
| DeepSWE v1.1 | 48.2% | 42.2% (Qwen3.8-27B) |
| SWE-bench Pro v2 Hard | 77.9% | 73.8% |

The Looped Transformer Architecture: How the Model Reasons Better
The secret behind the impressive performance of the Inkling 2 Mini is not just its compact size, but its innovative architecture: the Looped Transformer. According to Thinking Machines’ technical documentation, the model uses a Mixture of Experts (MoE) architecture with 24 layers, totaling 26.4 billion parameters, but with only 3.4 billion active parameters.
What makes this architecture particularly advanced is the way it processes information. Instead of passing each token through every transformer layer only once, the model executes its 12 central layers three times, while the initial and final six layers operate normally. This mechanism allows the model to perform 48 layer applications per token, even while storing weights for only 24 layers.
Practically speaking, this means the model can increase its effective active parameter count to 5.4 billion per token. This technique allows the Inkling 2 Mini to dedicate more internal reasoning “time” before generating the next token, drastically improving reasoning capability without proportionally increasing memory requirements.
📊 Looped Transformer Parameter Comparison
The Thinking Machines Context: From OpenAI to Accessible Models
Thinking Machines Lab is an American artificial intelligence startup, founded by Mira Murati, former CTO of OpenAI. The company has strategically positioned itself in the market, aiming to create open-source models that democratize access to advanced AI. The Inkling 2 Mini is a direct reflection of this mission: offering high performance in accessible formats.
The model was trained on the NVIDIA GB300 Grace Blackwell Ultra platform, leveraging its NVFP4 capabilities to optimize training efficiency. Most of the weights were trained using 4-bit NVFP4 precision, and the last four layers utilized 8-bit MXFP8 precision. Thinking Machines emphasized that because the weights are deployed in the same precision used during training, the model avoids additional conversion losses associated with post-training quantization.
This approach to hardware and software optimization reflects the industry trend of moving computational power from the cloud to the local device, making powerful models viable for everyday use on more modest laptops and workstations.

Impact for Developers: Advanced AI Without Cloud Dependency
For developers who want to create AI agents capable of running locally, the Inkling 2 Mini is a revolutionary tool. The ability to run such a capable model without depending on paid and slow cloud services eliminates significant barriers in the development of complex AI applications.
Despite the benefits, the Looped Transformer architecture raises important questions about security and transparency. Because the model can perform additional internal calculations without necessarily writing every step of reasoning, researchers warned that its behavior might become harder to monitor. Thinking Machines itself investigated this, noting that increasing internal passes raised the model’s ability to solve problems without producing a written chain of reasoning, going from 16 seconds to 37 seconds. However, researchers concluded that the model’s written reasoning remained comprehensible, without a significant reduction in monitorability at operational scales.
The model is available under the Apache 2.0 license, with weights for download on Hugging Face and local deployment support through Ollama. This solidifies the Inkling 2 Mini as a milestone in making cutting-edge artificial intelligence not just a feature of tech giants, but an accessible tool for the global community of developers and technology enthusiasts.
(Note: The technical information and data presented are based on the Thinking Machines Lab’s disclosure and do not represent a universal performance analysis, as the differences are in the configurational tests.)




