88Z AI, 21 November — Thinking Machines has launched Inkling 2 Mini, an open-weight looped transformer model with 26 billion parameters. By looping its middle layers, the model achieves deep reasoning with only 5.4 billion active parameters, scoring 20.1 on the AA Index. It is optimized for agentic tasks and coding, runs locally on Macs with 32GB RAM via Ollama, and maintains monitorability despite its recurrent architecture.
At a glance
- Inkling 2 Mini uses a 26B parameter looped transformer architecture, looping the middle 12 layers three times with only 5.4B active parameters.
- The model runs locally on Macs with 32GB RAM via Ollama, optimized for agentic tasks and coding.
- It costs 47% less per task than Inkling 2 Small on ARC-AGI-2 but scores 32 percentage points lower, with an AA Index score of 20.1.
- Research indicates that looping does not significantly reduce monitorability, as chains of thought remain legible despite being shorter.

Core Architecture and Parameter Scale
Responding to community feedback that the 594-billion-parameter Inkling 2 Small was too large, Thinking Machines has released Inkling 2 Mini. This open-weight model features 26 billion total parameters using a looped transformer architecture. It consists of 24 layers, where the first and last six layers run once, while the middle 12 layers loop three times. This design ensures that only 5.4 billion parameters are active per token generation. By enabling the model to "reason more while writing less," this looping mechanism enhances reasoning depth within a compact footprint, making it capable of handling agentic workloads and achieving a score of 20.1 on the AA Index.
Performance Metrics and Cost Efficiency
On the ARC-AGI benchmarks, Inkling 2 Mini demonstrates specific performance trade-offs. It scored 49.3% on ARC-AGI-2 (Verified) at a cost of $0.14 per task, and 84.5% on ARC-AGI-1 at $0.06 per task. Compared to Inkling 2 Small, Inkling 2 Mini reduces the cost per task on ARC-AGI-2 by 47%, although its score is 32 percentage points lower. This indicates that the model offers a cost-effective alternative for applications requiring moderate reasoning capabilities without the heavy computational burden of larger models, balancing performance with significant operational savings.
Local Deployment and Developer Ecosystem
To facilitate broad adoption, Thinking Machines has made Inkling 2 Mini highly accessible. The model is available on Ollama (pre-release version 0.40.17), allowing users to run it locally on Macs with 32GB or more of RAM. It also supports NVIDIA DGX Spark hardware and is integrated into the Tinker platform for both inference and fine-tuning, with immediate access via the Tinker Playground. Additionally, the company released a variable-loop research model to encourage the community to conduct their own experiments, exploring how different loop counts impact model behavior and performance.
Monitorability and Safety Research
The use of looping architectures raises concerns about the monitorability of internal model states. Thinking Machines addressed this in their release study. They found that no-Chain-of-Thought (no-CoT) capability improves with more passes. At this scale, while the chain of thought becomes shorter, it remains legible. Crucially, the team observed no meaningful drop in monitorability, ensuring that developers can still track and understand the model's reasoning process. This finding suggests that the benefits of deeper reasoning via looping can be achieved without significantly compromising transparency or safety oversight.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- @thinkymachinesSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.

