Thinking Machines Unveils Inkling 2 Mini Looped MoE Model
Thinking Machines has released Inkling 2 Mini, a 26.4B parameter model capable of running on a 32GB Mac with high performance on coding benchmarks.
Source: MarkTechPostThinking Machines has introduced Inkling 2 Mini, a new model designed to deliver high-level reasoning and coding capabilities on consumer-grade hardware. This release marks a significant step in making large-scale Mixture of Experts (MoE) architectures accessible to users with limited computational resources.
What is the Inkling 2 Mini architecture?
According to MarkTechPost, Inkling 2 Mini is a 26.4B-parameter transformer model that utilizes a looped Mixture of Experts (MoE) design. While the total parameter count sits at 26.4B, the model is optimized so that only 3.4B parameters are active during any single inference step. This efficiency is a hallmark of MoE architectures, which allow a model to possess a vast amount of knowledge while maintaining the computational speed of a much smaller network.
The "looped" aspect of the architecture suggests a specialized way of processing information through the transformer layers, aiming to maximize the utility of the parameters available within a constrained memory footprint.
How much hardware is required to run it?
One of the most notable aspects of the Inkling 2 Mini release is its hardware accessibility. The developers have optimized the model to run on a single 32GB Mac. For those with access to enterprise-grade hardware, the model can also be deployed on a single DGX Spark.
By targeting the 32GB RAM threshold, Thinking Machines is addressing a critical need for developers and researchers who wish to run sophisticated models locally without relying entirely on cloud-based API providers. This capability allows for greater privacy and lower latency for local development workflows.
How does the model perform on coding tasks?
Performance metrics suggest that the model is particularly well-suited for software engineering tasks. According to MarkTechPost, Inkling 2 Mini achieved a score of 79.2% on the SWE-bench Verified benchmark.
SWE-bench is a standard used to evaluate a model's ability to resolve real-world software engineering issues by interacting with actual code repositories. A score of nearly 80% indicates that the model possesses a high degree of proficiency in understanding complex codebases, identifying bugs, and implementing functional solutions, placing it in a competitive position for automated coding assistance.
Why does this matter for the AI ecosystem?
The release of Inkling 2 Mini highlights a growing trend in the industry: the push toward "dense intelligence" in smaller, more efficient packages. As models grow in total parameter count, the ability to selectively activate only a fraction of those parameters—as seen in this 3.4B active parameter configuration—is becoming essential for practical deployment.
For researchers, this provides a new tool for studying how looped architectures and MoE configurations can bridge the gap between massive cloud models and local edge devices. For developers, it offers a high-performing alternative for local IDE integration and automated testing environments.
How can you apply this to your workflow?
While Inkling 2 Mini is designed for local execution, the demand for high-performing coding models is a core focus for modern development environments. If you are looking to leverage advanced AI for programming, you can explore Grafy Code, our dedicated AI coding studio.
While we provide access to a variety of powerful models via the cloud, the principles behind Inkling 2 Mini—efficiency, specialized task performance, and high-reasoning capabilities—are exactly what developers look for when building complex software. For those interested in comparing how different models approach similar coding logic, you can also use Grafy Chat to run multiple models side by side.
Frequently asked questions
What is the total parameter count of Inkling 2 Mini?
The total parameter count for the Inkling 2 Mini model is 26.4B. However, due to its Mixture of Experts architecture, it only utilizes 3.4B active parameters during inference, which significantly reduces the computational load required to generate responses.
Can Inkling 2 Mini run on a standard laptop?
Yes, the model is designed to be highly efficient and can run on a Mac equipped with 32GB of RAM. This makes it possible for many developers to run a high-performing, 26.4B-parameter model locally without needing specialized server hardware.
How well does Inkling 2 Mini handle software engineering tasks?
Inkling 2 Mini performs strongly on software engineering benchmarks, specifically scoring 79.2% on the SWE-bench Verified test. This score demonstrates the model's capability to handle complex, real-world coding challenges and software repository interactions effectively.
Source: MarkTechPost — Meet Thinking Machines Inkling 2 Mini: 26B Looped MoE Runs on a 32GB Mac


