High operational expenses for cloud inference are forcing a structural shift toward local deployment architectures that bypass recurring token bills. An AI research venture has launched an open looped transformer model that reconfigures the compute path per token. Thinking Machines released Inkling 2 Mini, an open mixture-of-experts model featuring 26.4 billion parameters with 5.4 billion active per token because of internal loops.
Traditional architectures scale computational depth by increasing parameter volume, which directly raises hardware memory requirements. Inkling 2 Mini instead runs 12 of its 24 layers three times in a row for each token. As Thinking Machines explained, running these layers sequentially allows the system to reason internally for longer without generating additional text, arriving at answers with fewer output tokens.
Inkling 2 Mini runs 12 of its 24 layers three times in a row to reason internally without generating excess text.
The system processes text and images, incorporates a 1-million-token context window, and is distributed under an Apache 2.0 license.
Benchmark Metrics and Edge Deployment
Independent evaluations demonstrate that looped execution alters the cost equation for automated programming tasks. According to benchmark data, Inkling 2 Mini scored 48.2 percent on DeepSWE v1.1 and 46.5 percent on the Vals Index v2.2, leading open models under 100 billion parameters on agentic coding tests. In an evaluation conducted by Artificial Analysis, the model was awarded 20 points on its Intelligence Index.
This benchmark rating ties Inkling 2 Mini with Qwen3.8 27B while operating at roughly six times lower cost and four times the speed, per the Artificial Analysis findings. The performance shift allows engineering teams to deploy agentic coding workflows without committing capital to enterprise cloud endpoints. The model runs locally on Macs equipped with at least 32 GB of memory as well as on Nvidia DGX Spark hardware. Weights alongside a research version supporting variable loops are distributed on Hugging Face, while interactive testing is hosted in the Tinker Playground.
Corporate buyers evaluating on-premises deployment still face practical trade-offs regarding memory bandwidth and execution overhead on localized workstations. Sustained execution across repeated layer loops concentrates thermal load and memory access latency on client-side silicon rather than distributed cloud clusters. Enterprises must assess whether local workstation reliability offsets the infrastructure maintenance costs required when moving reasoning workloads out of centralized cloud environments.
How the picture shifted
In mid-July, the original Inkling relied on 975 billion parameters to claim open-weight capability, whereas Inkling 2 Mini routes tokens through repeating internal layers to keep demands modest. 5.4 billion active parameters match heavy enterprise benchmarks directly on client hardware.
Our stories: Mira Murati’s Inkling: The Most Powerful Open-Weights AI Model (July 16).