Thinking Machines Challenges Cloud-Based AI Paradigm with High-Performance "Looped" Inkling 2 Mini

In a significant move that could reshape the deployment landscape for local artificial intelligence, Thinking Machines Lab—the San Francisco-based startup helmed by former OpenAI chief technology officer Mira Murati—has unveiled Inkling 2 Mini. The new model is an open-weights, multimodal system engineered specifically to function offline on consumer-grade hardware, such as a 32GB MacBook Pro or a single desktop graphics card. By prioritizing architectural efficiency over sheer parameter scaling, Thinking Machines is positioning Inkling 2 Mini as a direct competitor to much larger, cloud-dependent frontier models, particularly in the fields of software engineering and agentic tool use.

At 26.4 billion total parameters and requiring approximately 20GB of memory, the model represents a departure from the "bigger is better" ethos that has dominated AI development for the past several years. Instead of simply increasing the size of the neural network, the engineering team at Thinking Machines has implemented a "looped transformer" architecture. This design enables the model to cycle data through its middle layers multiple times for each token generated, effectively squeezing deeper reasoning capabilities out of a compact hardware footprint. The result is a U.S.-built, Apache 2.0-licensed model that performs on par with dense 27B systems like those from Alibaba, while operating at significantly higher speeds and lower costs.

The Engineering Logic: How Looping Enhances Reasoning

To understand the innovation behind Inkling 2 Mini, one must first look at the limitations of standard transformer architectures. Traditional models process information in a linear, single-pass fashion from the first layer to the last. When a complex task requires more sequential reasoning than the model’s depth permits, the system typically compensates by generating long, verbose chains of thought. While effective, this approach is inefficient; it consumes excessive output tokens, increases latency, and drives up inference costs.

Inkling 2 Mini avoids these pitfalls through a sparse Mixture-of-Experts (MoE) design. It functions as a 24-layer, decoder-only model that routes each token to 8 of its 128 experts, activating roughly 3.4 billion parameters per pass. The architectural breakthrough lies in its execution: the first six layers, known as the prelude, and the final six layers run once. However, the middle 12 layers are cycled through three times using shared weights. This process mimics a human engineer who pauses to double-check a complex calculation multiple times before committing to a final answer, rather than creating extensive, messy "scratch work" that clutters the output.

These three passes result in 48 effective layer applications per token, providing the model with 5.4 billion effective active parameters without increasing its storage requirements. To ensure that information remains coherent across these iterations, Thinking Machines implemented attention residuals. These residuals allow each pass to read a gated "message" from the middle layers of the previous iteration, alongside the output from the initial prelude. Internal scaling tests conducted by the company demonstrated that a non-looping baseline model would require 140% more compute to match the reasoning performance of the looped architecture on large agentic traces.

To ensure the model remains practical for consumer silicon, the team pruned two-thirds of the experts mid-training—reducing the count from 384 to 128—and "healed" the network by distilling one trillion tokens from a larger internal Inkling model. This was followed by rigorous supervised fine-tuning and the application of 2.2 trillion tokens of reinforcement learning. By deploying the first 20 layers in 4-bit NVFP4 precision and the final four in 8-bit MXFP8 precision, the developers avoided the common accuracy degradation associated with post-training quantization, ensuring the model remains precise even on local machines.

Safety and Interpretability in a Looped System

The transition to a looped architecture raises valid questions regarding model oversight. If a system performs more reasoning internally rather than explicitly writing it down, does it become an opaque "black box"? Thinking Machines is proactively addressing these concerns. Alongside the release of the model, they have provided a variable-loop checkpoint—the TML-d24-varloop-research model—which allows researchers to test performance across one to five passes.

Furthermore, the company has partnered with the National Deep Inference Fabric and the research firm Goodfire to study the mechanistic interpretability of the model. Preliminary monitorability tests have shown that looping extends the model’s "no-CoT" reasoning horizon from 16 seconds to 37 seconds, while simultaneously reducing the length of chain-of-thought outputs by more than half. Crucially, these traces remain human-legible, and the model has shown no indication of increased risk regarding covert sabotage on the LinuxArena control benchmark, suggesting that efficiency does not come at the cost of safety.

Benchmarks and Real-World Performance

Inkling 2 Mini is natively multimodal, capable of processing both UTF-8 text and images within a 1-million-token context window. The model’s real-world utility was demonstrated in offline scenarios, such as running in airplane mode on a 48GB M5 Pro MacBook Pro to build a functional flight simulator in minutes, or managing 20 persistent autonomous agents in a local simulation.

On standardized benchmarks, the model consistently outperforms other sub-100B open-weights systems. It leads in agentic coding and tool use, achieving 48.2% on DeepSWE v1.1, 77.9% on SWE-bench Pro v2 Hard, and 46.5% on the Vals Index v2.2. It also scored 85.3% on DeepSearchQA. On the Stanford intelligence-per-joule benchmark, Inkling 2 Mini achieved a score of 0.703, placing it well ahead of peers like the Qwen3.6-35B-A3B and Gemma-4-26B-A4B.

However, Thinking Machines is transparent about the model’s limitations. Because factual recall is primarily a function of total parameter count, the model trails denser competitors like Qwen3.8-27B in pure knowledge retrieval. It also shows occasional performance gaps in multi-turn chat and complex constraint satisfaction, and some users have noted minor issues with clipped phrasing in non-English languages. Despite this, the economic advantage is stark: when compared to the Qwen3.8-27B, the Inkling 2 Mini uses roughly 2.8 times less compute, runs four times faster, and costs six times less per task at first-party API rates.

Enterprise Adoption and Licensing

By releasing the model under the Apache 2.0 license, Thinking Machines has made a strategic choice to lower the barrier for enterprise adoption. This permissive license allows for the commercial use, modification, and redistribution of the model without the need for complex royalty structures or the requirement to open-source downstream applications. For legal and procurement teams at large organizations, this provides a clear, reliable path toward deploying private, on-premises AI solutions that do not rely on third-party cloud providers. This approach avoids the commercial friction found in other licensing models, such as those that impose revenue thresholds or usage restrictions on service providers.

Community Reception

The developer community has responded with significant enthusiasm. Within 24 hours of its release, the model saw rapid integration into local inference runtimes. Infrastructure partners like UnslothAI were quick to provide GGUF builds and llama.cpp support, emphasizing the model’s ability to run on standard hardware with as little as 16GB of RAM. Apple silicon specialists also reported high throughput figures, noting that the looping architecture is particularly well-suited for the memory-constrained environments of consumer laptops.

Independent testers have echoed these sentiments, with many highlighting the model’s ability to handle creative and technical tasks that typically overwhelm smaller systems. While a small subset of the community has pointed to the need for further refinement in areas like game-mechanics generation and specific language quirks, the consensus remains that Inkling 2 Mini is a landmark release.

A Rapid March Toward Local Intelligence

The launch of Inkling 2 Mini is the latest chapter in a remarkably fast trajectory for Thinking Machines Lab. Founded in early 2025 by Mira Murati, John Schulman, and a cohort of veteran researchers, the startup secured a $2 billion seed round just months later. Since its debut product, the Tinker fine-tuning API, the company has consistently pushed the boundaries of multimodal interaction and open-science foundation models.

Following the releases of the 975B-parameter Inkling, the 276B-parameter Inkling-Small, and the 594B-parameter Inkling 2 Small, this latest iteration finally fulfills the long-standing promise of truly high-agency, local AI. By rethinking the fundamental way a transformer allocates its compute, Thinking Machines has moved the needle for developers and enterprises alike. For those seeking to run high-agency coding assistants on a laptop or a single workstation, the era of the looped transformer has officially begun.

Share:

Laily UPN writes for Tech Maze.